An Overview of Existing Applications of Artificial Intelligence in Histopathological Diagnostics of Lymphoma: A Scoping Review

1. Introduction to Lymphomas and Their Diagnostic Process

Our previous review, published recently in Electronics [1] was devoted to analysing methods of artificial intelligence used in the histopathological diagnostics of leukemias. This time we decided to devote a similar work to lymphomas - a very heterogenous group of haematological neoplasms. While leukemias are neoplasms originating from intermediate stages of lines of haematopoiesis (myeloid, lymphoid), lymphomas originate from mature forms or from developmental forms of lymphoid line - lymphocytes T & lymphocytes B. These cells are called lymphatic cells and within the body they reside mainly in lymphnodes, spleen, thymus and bone marrow. An example of such a lymphoma can be primary cerebral lymphoma, frequently occuring in immunocompromised patients suffering from AIDS [2]. The symptomatology of lymphomas can differ depending on the location of disease, but often involves swelling of lymphnodes (unspecific, can occur also in e.g. infections). The so-called B-symptomes (fever, night sweats, and weight loss) are often present, yet they are highly unspecific and can also occur in numerous other neoplastic diseases. Because of the unspecific clinical presentation the diagnosis of lymphoma is often delayed and can take a long time before it is established.

There are in total over 50 entities that can be classified as lymphomas [3]. Historically, they have been classified into Hodgkin lymphoma and a very heterogenous group of non-Hodgkin lymphomas (NHL). While this classification is still used, it lost much on relevance when WHO introduced new classification that bases on the original subgroup of cells from which a lymphoma stems. Thus, we can speak of T-cell lymphomas and B-cell lymphomas. We can also divide lymphomas according to their clinical course - some of lymphomas are progressing very slowly and for a long time they don’t yield any clinical symptoms; these lymphomas are called indolent. Other lymphomas can progress very aggresively and quickly lead to the patients’ death. The diversity of lymphomas makes their classification quite difficult. The line of division between leukemias and lymphomas is not always clear, sometimes being a matter of convention only; some nosological units are classified into both subgroups of haematological malignancies, one of them being the SLL/CLL. These two terms are essentially the same neoplastic disease deriving from small B-cell lymphocytes, and the only difference is the localisation - when the disease is occuring mainly in bone marrow it is counted as leukemia and when it occurs in lymphnodes it is classified as lymphoma; in order to add to confusion one has to note that this type of lymphoma can also transform into another, more malignant type (a phenomenon called Richter transformation) [4]. Lymphomas can originate in almost every organ of the body, since lymphatic tissue is almost omnipresent. The staging also differs from other tumours - lymphomas are not routinely staged using the TNM Classification, but rather the AnnArbor scale, which takes into account the location of diseased lymphnodes, and the presence of absence of systemic symptoms. xplicit enumeration of lymphomas is outside the scope of this work, interested readers can find relevant information in appropriate publications [5]. The aforementioned facts serve to illustrate the heterogeneity and the convolution of lymphomas which have direct consequences on the effectivity and complexity of their diagnosis.

The diagnosis of lymphomas relies on histopathology in as much as the one of leukemias, yet there are significant differences in the diagnostic process. In order to diagnose lymphoma it is not enough to find malignant cells under the microscope, one also needs to examine the architecture of affected tissue. Thus, cytopathology is not very helpful and does not suffice to make a diagnosis of lymphoma. The standard method is an excision biopsy of affected lymph node. Lymphnode is then analysed using histological methods. In order to differentiate between lymphoma type ancillary techniques are used, including immunohistochemistry staining (different cells will stain differently, thus giving clues as to lymphoma origins). Flow cytometry, and genetic analyses are also used to secure the diagnosis. The majority of those ancillary methods fall outside the scope of our review - in our work we take into account only those models that work on histopathology slides (stained with routine stains, or with immunohistochemical ones). There are many possibilities for AI to help with the analysis of such histological slides, and this scoping review is intended to enumerate some of best performing models.

The methods used for digital image analysis basing on deep learning are almost as ubiquitous as the lymphomas. There are many models published in scientific literature, but not all of them have already been registered and used in a daily routine of clinical praxis. As for today, there are only 6 entries on the FDA list of approved AI-based solutions in the domain of pathology, none of these entries dealing with the problematics of lymphoma pathology. The description of the methods has already been done in our previous review.

2. Bibliometric Analysis

This scoping review was conducted and reported in accordance with the PRISMA Extension for Scoping Reviews (PRISMA-ScR) guidelines. No review protocol was registered for this scoping review, as protocol registration is not mandatory for scoping reviews, unlike systematic reviews which require PROSPERO registration.

2.1. Methodology:

The bibliometric analysis conducted within the framework of this review comprised a systematic search of three major scientific publication databases (Scopus, PubMed, Web of Science) to identify publications addressing the application of artificial intelligence in histopathological diagnosis of lymphomas. The search encompassed the period 2016–2025 and was restricted to English-language publications of the Article and Review types. The application of seven precisely formulated search queries enabled the identification of methodologically relevant publications, accounting for variations in study modality, algorithmic architectures, and disease-specific pathological terminology. Detailed inclusion and exclusion criteria, search strategy, and selection procedures are described in the subsequent subsections.

2.2. Queries:

Seven diversified search queries were formulated in Advanced Search mode, differentiated by modality (WSI vs. cytology), task type (diagnostics/classification vs. segmentation), keywords related to lymphomas (e.g., lymphoma, DLBCL, Hodgkin, MCL, FL), network architectures/models (CNN/ViT/transformers), as well as strictly pathological terminology (H&E, IHC, digital pathology). Each query was syntactically adapted to the specific requirements of the respective publication database.

Table 1. Queries.

id	Database	Query	Search Date	Number of records
Q1	Scopus, PubMed, Web of Science	(("artificial intelligence" OR "machine learning" OR "deep learning" OR AI) AND ("histopatholog*" OR "digital pathology" OR "whole slide imag*" OR WSI OR "microscopic image*" OR cytolog* OR "lymph node biopsy") AND (lymphoma OR "Hodgkin lymphoma" OR "non-Hodgkin lymphoma" OR DLBCL OR "diffuse large B-cell lymphoma" OR "follicular lymphoma" OR "mantle cell lymphoma" OR "marginal zone lymphoma" OR "Burkitt lymphoma" OR "peripheral T-cell lymphoma" OR "anaplastic large cell lymphoma") AND (diagnos* OR classif* OR "diagnostic support" OR "computer-aided diagnos*")) AND NOT TITLE-ABS-KEY(radiology OR radiomic* OR CT OR "computed tomography" OR MRI OR PET OR ultrasound OR "flow cytometr*" OR genomic* OR sequencing OR "gene expression" OR leukemia OR myeloma)	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus: 175; PubMed: 59; Web of Science: 82
Q2	Scopus, PubMed, Web of Science	TITLE-ABS-KEY(("deep learning" OR "machine learning" OR "artificial intelligence") AND (subtyp* OR "cell-of-origin" OR COO OR "double-hit" OR "triple-hit" OR "EBV-positive" OR "grade 3B") AND (DLBCL OR "diffuse large B-cell lymphoma" OR "follicular lymphoma" OR "mantle cell lymphoma" OR "Hodgkin lymphoma" OR "classical Hodgkin lymphoma" OR "nodular lymphocyte-predominant Hodgkin lymphoma" OR "T-cell lymphoma" OR "anaplastic large cell lymphoma") AND ("histopatholog*" OR "digital pathology" OR "whole slide imag*" OR WSI OR "H&E" OR "haematoxylin and eosin" OR "immunohistochemistry" OR IHC) AND (classif* OR differentiat* OR "risk stratification")) AND NOT TITLE-ABS-KEY(radiology OR CT OR MRI OR PET OR "flow cytometr*" OR genomic* OR sequencing OR leukemia OR myeloma)	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus:24; PubMed: 11; Web of Science: 18
O3	Scopus, PubMed, Web of Science	TITLE-ABS-KEY(("convolutional neural network" OR CNN OR "vision transformer" OR ViT OR "multiple instance learning" OR MIL OR "graph neural network" OR GNN) AND ("whole slide imag*" OR WSI OR histopatholog* OR "digital pathology") AND (segmentation OR "cell detection" OR "nuclei segmentation" OR "instance-level" OR "weakly supervised" OR attention OR patch OR tile) AND (lymphoma OR DLBCL OR "follicular lymphoma" OR "Hodgkin" OR "mantle cell lymphoma" OR "T-cell lymphoma")) AND NOT TITLE-ABS-KEY(radiology OR CT OR MRI OR PET OR "flow cytometr*" OR genomic* OR leukemia OR myeloma)	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus:44; PubMed: 8; Web of Science: 17
Q4	Scopus, PubMed, Web of Science	TITLE-ABS-KEY(("artificial intelligence" OR "machine learning" OR "deep learning") AND lymphoma AND ("digital pathology" OR histopatholog* OR "whole slide imag*") AND ("systematic review" OR "scoping review" OR "meta-analysis" OR bibliometric OR "state of the art")) AND NOT TITLE-ABS-KEY(radiology OR CT OR MRI OR PET OR "flow cytometr*" OR genomic* OR leukemia OR myeloma)	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus:17; PubMed: 5; Web of Science: 9
Q5	Scopus, PubMed, Web of Science	TITLE-ABS-KEY(("machine learning" OR "deep learning" OR "artificial intelligence") AND (IHC OR "immunohistochemistry" OR "H&E" OR "haematoxylin and eosin" OR "multiplex immunofluorescence" OR "tissue microarray" OR TMA) AND (feature* OR morphometric OR "texture" OR "nuclei" OR "cell segmentation" OR "tumor microenvironment" OR TME OR "germinal center") AND (lymphoma OR DLBCL OR "follicular lymphoma" OR "Hodgkin lymphoma" OR "mantle cell lymphoma") AND (diagnos* OR classif*)) AND NOT TITLE-ABS-KEY(radiology OR CT OR MRI OR PET OR "flow cytometr*" OR genomic* OR sequencing OR leukemia OR myeloma)	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus:50; PubMed: 17; Web of Science: 25
Q6	Scopus, PubMed, Web of Science	TITLE-ABS-KEY(("artificial intelligence" OR "machine learning" OR "deep learning") AND (lymphoma) AND ("histopathology" OR "digital pathology" OR "whole slide imag*" OR WSI OR cytolog*))	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus:241; PubMed: 77; Web of Science: 98
Q7	Scopus, PubMed, Web of Science	TITLE-ABS-KEY(("artificial intelligence" OR "machine learning" OR "deep learning") AND (DLBCL OR "diffuse large B-cell lymphoma" OR "follicular lymphoma" OR "Hodgkin lymphoma" OR "mantle cell lymphoma") AND ("histopathology" OR "whole slide imag*" OR WSI OR IHC OR "H&E"))	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus:71; PubMed: 36; Web of Science: 45

Table 1. Queries.

id	Database	Query	Search Date	Number of records
Q1	Scopus, PubMed, Web of Science	(("artificial intelligence" OR "machine learning" OR "deep learning" OR AI) AND ("histopatholog*" OR "digital pathology" OR "whole slide imag*" OR WSI OR "microscopic image*" OR cytolog* OR "lymph node biopsy") AND (lymphoma OR "Hodgkin lymphoma" OR "non-Hodgkin lymphoma" OR DLBCL OR "diffuse large B-cell lymphoma" OR "follicular lymphoma" OR "mantle cell lymphoma" OR "marginal zone lymphoma" OR "Burkitt lymphoma" OR "peripheral T-cell lymphoma" OR "anaplastic large cell lymphoma") AND (diagnos* OR classif* OR "diagnostic support" OR "computer-aided diagnos*")) AND NOT TITLE-ABS-KEY(radiology OR radiomic* OR CT OR "computed tomography" OR MRI OR PET OR ultrasound OR "flow cytometr*" OR genomic* OR sequencing OR "gene expression" OR leukemia OR myeloma)	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus: 175; PubMed: 59; Web of Science: 82
Q2	Scopus, PubMed, Web of Science	TITLE-ABS-KEY(("deep learning" OR "machine learning" OR "artificial intelligence") AND (subtyp* OR "cell-of-origin" OR COO OR "double-hit" OR "triple-hit" OR "EBV-positive" OR "grade 3B") AND (DLBCL OR "diffuse large B-cell lymphoma" OR "follicular lymphoma" OR "mantle cell lymphoma" OR "Hodgkin lymphoma" OR "classical Hodgkin lymphoma" OR "nodular lymphocyte-predominant Hodgkin lymphoma" OR "T-cell lymphoma" OR "anaplastic large cell lymphoma") AND ("histopatholog*" OR "digital pathology" OR "whole slide imag*" OR WSI OR "H&E" OR "haematoxylin and eosin" OR "immunohistochemistry" OR IHC) AND (classif* OR differentiat* OR "risk stratification")) AND NOT TITLE-ABS-KEY(radiology OR CT OR MRI OR PET OR "flow cytometr*" OR genomic* OR sequencing OR leukemia OR myeloma)	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus:24; PubMed: 11; Web of Science: 18
O3	Scopus, PubMed, Web of Science	TITLE-ABS-KEY(("convolutional neural network" OR CNN OR "vision transformer" OR ViT OR "multiple instance learning" OR MIL OR "graph neural network" OR GNN) AND ("whole slide imag*" OR WSI OR histopatholog* OR "digital pathology") AND (segmentation OR "cell detection" OR "nuclei segmentation" OR "instance-level" OR "weakly supervised" OR attention OR patch OR tile) AND (lymphoma OR DLBCL OR "follicular lymphoma" OR "Hodgkin" OR "mantle cell lymphoma" OR "T-cell lymphoma")) AND NOT TITLE-ABS-KEY(radiology OR CT OR MRI OR PET OR "flow cytometr*" OR genomic* OR leukemia OR myeloma)	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus:44; PubMed: 8; Web of Science: 17
Q4	Scopus, PubMed, Web of Science	TITLE-ABS-KEY(("artificial intelligence" OR "machine learning" OR "deep learning") AND lymphoma AND ("digital pathology" OR histopatholog* OR "whole slide imag*") AND ("systematic review" OR "scoping review" OR "meta-analysis" OR bibliometric OR "state of the art")) AND NOT TITLE-ABS-KEY(radiology OR CT OR MRI OR PET OR "flow cytometr*" OR genomic* OR leukemia OR myeloma)	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus:17; PubMed: 5; Web of Science: 9
Q5	Scopus, PubMed, Web of Science	TITLE-ABS-KEY(("machine learning" OR "deep learning" OR "artificial intelligence") AND (IHC OR "immunohistochemistry" OR "H&E" OR "haematoxylin and eosin" OR "multiplex immunofluorescence" OR "tissue microarray" OR TMA) AND (feature* OR morphometric OR "texture" OR "nuclei" OR "cell segmentation" OR "tumor microenvironment" OR TME OR "germinal center") AND (lymphoma OR DLBCL OR "follicular lymphoma" OR "Hodgkin lymphoma" OR "mantle cell lymphoma") AND (diagnos* OR classif*)) AND NOT TITLE-ABS-KEY(radiology OR CT OR MRI OR PET OR "flow cytometr*" OR genomic* OR sequencing OR leukemia OR myeloma)	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus:50; PubMed: 17; Web of Science: 25
Q6	Scopus, PubMed, Web of Science	TITLE-ABS-KEY(("artificial intelligence" OR "machine learning" OR "deep learning") AND (lymphoma) AND ("histopathology" OR "digital pathology" OR "whole slide imag*" OR WSI OR cytolog*))	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus:241; PubMed: 77; Web of Science: 98
Q7	Scopus, PubMed, Web of Science	TITLE-ABS-KEY(("artificial intelligence" OR "machine learning" OR "deep learning") AND (DLBCL OR "diffuse large B-cell lymphoma" OR "follicular lymphoma" OR "Hodgkin lymphoma" OR "mantle cell lymphoma") AND ("histopathology" OR "whole slide imag*" OR WSI OR IHC OR "H&E"))	Scopus: 13 July 2025; PubMed: 9 October 2025; Web of Science: 26 October 2025	Scopus:71; PubMed: 36; Web of Science: 45

2.3. Screening Protocol:

Records retrieved from all databases were combined into a unified dataset. Title and abstract screening was conducted independently by two reviewers using standardized forms based on pre-defined inclusion/exclusion criteria. Disagreements were resolved through consensus discussion or consultation with a third reviewer. A two-stage screening approach was employed:

1.

initial automated filtering based on clearly ineligible publication types, languages, and time frame;

2.

manual review of titles and abstracts against detailed eligibility criteria.

Full-text articles flagged as potentially eligible were retained for data extraction phase.

2.4. Inclusion/Exclusion Criteria

Table 2. Inclusion/Exclusion Criteria.

Category	Inclusion	Exclusion
Data type	Images (WSI, H&E, IHC, cytology)	Non-imaging data (omics, RNA-seq, flow cytometry)
Language	English (optionally Polish)	Other languages
Publication type	Articles, Reviews	Other (editorials, books, conference abstracts, preprints)
Population	Human samples	Animal models, in vitro samples
Time frame	2016–2025	Publications before 2016
Algorithm type	Deep learning, machine learning, CNN, ViT, transformers, MIL	Purely statistical approaches without deep learning componen

Table 2. Inclusion/Exclusion Criteria.

Category	Inclusion	Exclusion
Data type	Images (WSI, H&E, IHC, cytology)	Non-imaging data (omics, RNA-seq, flow cytometry)
Language	English (optionally Polish)	Other languages
Publication type	Articles, Reviews	Other (editorials, books, conference abstracts, preprints)
Population	Human samples	Animal models, in vitro samples
Time frame	2016–2025	Publications before 2016
Algorithm type	Deep learning, machine learning, CNN, ViT, transformers, MIL	Purely statistical approaches without deep learning componen

2.5. Deduplication:

The combined search returned 612 records. After title and abstract screening (314 records retained) and systematic deduplication by DOI and title matching, 182 unique publications were included in the final dataset.

2.6. Data Extraction and Charting:

Data extracted from 182 publications included: publication metadata (authors, year, journal, country, affiliation), citation counts, and keywords. Bibliometric indicators were synthesized to examine temporal trends, geographical distribution, institutional productivity, and thematic profiles. Results were visualized through charts showing publications/citations per year, country contributions, leading affiliations, and top 15 keywords.

2.7. Results

The bibliometric analysis identified 182 publications meeting the inclusion criteria, confirming a dynamic and expanding research landscape in the application of artificial intelligence to lymphoma histopathology. The following sections present the key findings across temporal, thematic, and editorial dimensions.

Temporal Trends in Publication Output Figure 1 illustrates the publication dynamics across the study period (2016-2025). The data reveal a pronounced upward trajectory, particularly from 2022 onwards, indicating accelerating research activity in recent years. The 2024-2025 period shows sustained publication volume, suggesting that AI applications in lymphoma diagnostics remain an active and evolving area of investigation. This temporal pattern reflects both the maturation of deep learning methodologies and growing clinical interest in computational pathology solutions.

The 182 included publications were distributed across diverse publication outlets, reflecting the interdisciplinary nature of the field. Figure 2 presents the top 10 journals by publication count. Diagnostics and Frontiers in Oncology emerged as the leading venues with 9 publications each, followed by Cancers with 8 publications. Other prominent outlets include Computers in Biology and Medicine, IEEE Journal of Biomedical and Health Informatics, and Journal of Pathology Informatics, each contributing 4 publications. This distribution underscores the multidisciplinary appeal of the topic, spanning clinical pathology, biomedical informatics, oncology, and computer science domains.

Thematic Profile and Methodological Focus Figure 3 displays the 10 most frequently occurring keywords across the included publications. Deep Learning” emerged as the dominant keyword (56 occurrences), followed by Lymphoma” (36 occurrences) and Machine Learning” and Artificial Intelligence” (both 35 occurrences). Other prominent keywords include Digital Pathology” (17 occurrences), Convolutional Neural Network” (11 occurrences), and Diffuse Large B-Cell Lymphoma” (10 occurrences). The prevalence of these terms confirms the central role of neural network-based approaches and highlights the focus on specific lymphoma subtypes. Keywords such as Histopathology”, Immunohistochemistry”, and Classification” further reinforce the morphological and diagnostic orientation of the research corpus.

2.8. Limitations

This bibliometric analysis, although conducted in accordance with PRISMA-ScR guidelines, is subject to several important methodological limitations that should be considered when interpreting the results.

First, the scope of databases was restricted to three major indexing sources (Scopus, PubMed, Web of Science), which may have resulted in the omission of significant publications indexed exclusively in other databases (e.g., IEEE Xplore, Google Scholar) or available in preprint repositories (arXiv, bioRxiv). Differences in journal coverage across databases and variations in indexing policies may introduce systematic selection bias, particularly with respect to publications from non-English-speaking countries or open-access journals with limited reach.

Second, language restrictions (English-language publications only) may have led to underrepresentation of research activity in regions where local-language publishing is preferred, such as China, Japan, and Germany. Given the significant contribution of Asian research institutions to the advancement of AI in pathology, this limitation may affect the geographical representativeness of the results.

Third, the search strategy, despite employing seven precisely formulated queries adapted to the specifics of each database, was characterized by high specificity at the potential cost of reduced sensitivity. Publications employing non-standard terminology, alternative acronyms (e.g., "virtual slides" instead of "WSI"), or emerging methodologies (e.g., foundation models, self-supervised learning) may have been overlooked. The intentional exclusion of publications addressing radiology, flow cytometry, and omics analyses, although justified by the scope of the review, narrows the perspective to morphological image-based diagnostics exclusively.

Fourth, the deduplication and screening process, based on automated title matching and manual abstract review, introduces an element of subjectivity, although screening was conducted independently by two reviewers. The absence of full-text assessment for all potentially qualifying publications may have resulted in both false-positive and false-negative inclusions. Additionally, publications with unclear or very brief abstracts may have been misclassified.

Fifth, bibliometric metrics (citation counts, institutional productivity, journal distribution) reflect only research visibility and activity, not methodological quality, result reliability, or potential clinical impact. Citation counts recorded at the time of data extraction are dynamic and subject to significant changes over time. No assessment of risk of bias for individual studies was conducted, which is typical for bibliometric reviews but represents a limitation in the context of evaluating evidence quality.

Finally, the time frame (2016–2025) may not fully capture the most recent methodological trends, particularly if 2025 publications were incompletely indexed at the time of data extraction. Moreover, the rapid development of AI technologies, including the emergence of foundation models and multimodal architectures in recent years, may not be fully reflected in the publication corpus.

2.9. Conclusions

This bibliometric analysis identified 182 publications meeting the defined inclusion criteria, documenting a dynamic increase in research activity in the field of artificial intelligence applications in histopathological diagnosis of lymphomas. The results confirm a marked intensification of publications beginning in 2022, correlating with the maturation of deep learning architectures and growing clinical interest in digital pathology solutions.

Analysis of journal distribution reveals the interdisciplinary nature of the research, encompassing both clinical journals (Diagnostics, Frontiers in Oncology, Cancers) and biomedical and informatics-focused publications (IEEE Journal of Biomedical and Health Informatics, Journal of Pathology Informatics, Computers in Biology and Medicine). The presence of publications across leading journals from diverse domains confirms the acceptance of AI methods as a legitimate diagnostic tool in hematopatology.

The keyword profile unequivocally indicates the dominance of deep learning approaches, with particular emphasis on convolutional neural networks (CNNs) and classification tasks. The frequent occurrence of terms related to digital pathology (Digital Pathology, Histopathology, Immunohistochemistry) confirms the morphological orientation of the research, while the presence of specific lymphoma subtype designations (e.g., DLBCL) indicates the clinical specialization of the publications.

The bibliometric findings underscore the field’s transition from an exploratory phase (proof-of-concept) toward consolidation, as evidenced by the growing number of validation-focused publications and the increasing emphasis on diagnostic accuracy. The research landscape demonstrates robust interdisciplinary engagement and sustained institutional commitment to advancing AI-assisted lymphoma diagnostics. This comprehensive mapping of the publication landscape provides a foundation for identifying gaps in knowledge, recognizing leading research centers, and understanding the temporal and thematic evolution of AI applications in histopathology. These bibliometric insights will be further contextualized in the discussion section through detailed analysis of methodological approaches, emerging trends, and future research priorities in the field.

3. Image Processing Methods Used for Histopathological Diagnostics of Lymphomas

Based on the bibliometric analysis, 36 publications were identified as relevant to the use of artificial intelligence (AI) in lymphoma diagnostics. These studies represent different approaches to AI-based diagnostic support and were selected for detailed review. To provide a clear overview of current research directions and methodological differences, the applied models are summarized in Table 3. For each study, six categories of information are highlighted: the material used for diagnosis, the performed diagnosis defined by the detected classes, the use of the model, the dataset used for model training, and the obtained results. In most publications, the model is used for lymphoma diagnosis through classification. The performance of such models is evaluated using accuracy ACC, expressed by formula (1), which is an intuitive metric that indicates the percentage of correct decisions made by the model.

$$ACC={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(1)

Model accuracy, denoted as $ACC$, is calculated based on the confusion matrix, which divides the model’s results into four groups:

• True Positive (TP) — the model predicted true and it actually is true,
• True Negative (TN) — the model predicted false and it actually is false,
• False Positive (FP) — the model predicted true but it is false,
• False Negative (FN) — the model predicted false but it is true.

The confusion matrix is used to compute various performance metrics. One of these metrics is the true positive rate (TPR) known as sensitivity expressed by formula (2). It measures the proportion of actual positive cases that are correctly identified by the model. This metric is particularly important in medical applications, as it is preferable to refer a healthy individual for additional testing than to miss a sick one.

$$TPR={\displaystyle \frac{TP}{TP+FN}}$$

(2)

Another important metric derived from the confusion matrix is the false positive rate (FPR) expressed by formula (3). In medical applications, it represents the proportion of healthy cases incorrectly classified as sick by the model.

$$FPR={\displaystyle \frac{FP}{FP+TN}}$$

(3)

Specificity (SPC) is another metric derived from the confusion matrix. It measures the proportion of true negative cases correctly identified by the model, reflecting its ability to avoid false positives. In medical applications, high specificity ensures that healthy individuals are not subjected to unnecessary additional tests.

$$SPC={\displaystyle \frac{TN}{TN+FP}}$$

(4)

Classification models predict classes as probabilities in range from 0 to 1, with standard decision threshold of 0.5. Lowering this threshold below $0.5$ increases sensitivity (TPR), thereby reducing the risk of missed cases, but also rise FPR, causing a higher number of false alarms. In other hand, raising the threshold above $0.5$ reduces FPR but decreases TPR, results in fewer false alarms, at the cost of missing some positive cases.

The relationship between TPR and FPR across all possible decision thresholds is described by the Receiver Operating Characteristic (ROC) curve, where TPR is plotted on the Y-axis, while the FPR is shown on the X-axis. Each point on the curve corresponds to a different decision threshold. An ideal classifier passes through the upper-left corner (0,1), whereas a random model follows the diagonal. The diagnostic value of the ROC curve is quantified by the area under the ROC curve (AUC) ranges from 0.5, which corresponds to a random classifier following the diagonal, to 1.0, which indicates a perfect model.

The majority of models presented in Table 3 report performance using ACC and AUC, as shown in the Results column. These models detect lymphomas by classifying them into detailed classes, which are listed in the Diagnose column. A total of 26 classes are identified. In most cases, classification is performed using three selected classes. However, there are also publications presenting classifications with a larger number of classes, up to eight. Among the reviewed publications, seventeen classes appear only once, and only four classes appear more than three times. These four most frequently detected classes are diffuse large B-cell lymphoma (DLBCL) used in 13 publications with ACC up to 97% and AUC up to 96.9% [6,7,8,9,10,11,12,13,14,15,16,17,18], follicular lymphoma (FL) used in 11 publications with ACC and AUC up to 100% [8,11,19,20,21,22,23,24,25,26,27], lymphocytic leukemia (CLL) used in 9 publications with ACC and AUC up to 100% [15,18,19,20,22,25,26,27,28], and mantle cell lymphoma (MCL) detected in 7 publications with ACC and AUC up to 100% [11,20,21,22,25,26,27].

In the analyzed publications, classical artificial intelligence models with ACC and AUC up to 100% were applied [18,20,24,28,29,30,31,32,33] in particular Decision Tree (DT) [20,31], Random Forest (RF) [18,20], Support Vector Machine (SVM) [32], and Bayesian Neural Network (BNN) [24] models were identified.

In most publications, convolutional neural network (CNN) architectures were used. In four publications custom solutions were proposed with reported ACC up to 97% and AUC up to 96.9% [6,7,8,31]. On the other hand, 20 publications present models based on well-known and validated CNN architectures for image processing [9,11,12,14,15,19,22,23,25,26,27,32,33,34,35,36,37,38,39,40]. Among them, 11 models use the ResNet architecture with reported ACC and AUC up to 100% [9,11,14,19,23,27,32,34,35,36,37]. EfficientNet was used in four models with ACC up to 95.56% and AUC up to 87% [15,35,36,38]. The MobileNet, YOLO, VGG16, and DenseNet architectures are each used twice with reported ACC and AUC up to 100% [12,19,25,26,32,39,40,12]. GoogLeNet, AlexNet, U-Net, and SqueezeNet are each used once, with ACC reported by the authors as reaching up to 99.87% [22,26,33,35].

The Attention mechanism, originally known from the Transformer architecture for natural language processing, was also used for lymphoma detection in 4 models achieved ACC and AUC up to 96% [9,12,14,34]. There are also three models using Vision Transformer (ViT), with reported effectiveness reaching ACC up to 99.87% and AUC of 85.6% [13,17,22].

Table 3. Overview of existing AI models in lymphoma diagnostics.

Ref.	Material	Diagnose	Use of the Model	Dataset	AI architecture / method	Results
[34]	skin biopsies	MF, BIDs	Slide-level classification	private dataset of 924 H&E-stained whole-slide images from skin biopsies: 233 patients with early-stage MF and 353 patients with benign inflammatory dermatoses	clustering-constrained attention multiple instance learning (CLAM), ResNet50, CTransPath, UNI	ACC: CLAM-ResNet50: 0.918, CLAM-CTransPath: 0.921, CLAM-UNI: 0.971
[19]	Bone Marrow	LCT, FL, CLL	Slide-level classification	dataset of 71 patients comprising whole-slide images (WSI) of H&E-stained bone marrow biopsies: 21 patients with follicular lymphoma (FL) and 50 patients with chronic lymphocytic leukemia (CLL)	Convolutional Neural Network (CNN),ResNet-50, MobileNet	AUC: CNN(Ratio Pred)-(FL): 0.881, CNN(End-to-End)-(FL): 0.923, CNN(Ratio Pred)-(CLL): 0.780, CNN(End-to-End)-(CLL): 0.823
[6]	Hematoxylin and Eosin (H&E) stained whole-slide images of lymph nodes	BLN, DLBCL, BL, SLL	Classification into 4 diagnostic categories	Whole-slide images obtained from two public sources: (1) Virtual Pathology at the University of Leeds (355,966 WSIs, 114.92 TB), and (2) Virtual Slide Box from the University of Iowa (>1,000 WSIs). From these, 128 cases (32 per class) were selected.	Convolutional Neural Network (CNN) implemented in Python using TensorFlow and Keras.	ACC:CNN-(Image by Image):0.95, CNN(set-by-set classification using majority voting, 3/5 agreement):1
[35]	Hematoxylin and Eosin (H&E) stained whole-slide images of lymphoid tissue (lymph node and tonsil)	BL	Slide-level classification	160 patients total: 90 BL and 70 control (tonsillectomy or reactive lymphoid samples).	Multiple-Instance Learning (MIL),ResNet50, EfficientNet, and GoogLeNet (Inception-v1)	AUC:ResNet50: 0.90, ResNet50+Att: 0.91, GoogLeNet: 0.93, GoogLeNet+Att: 0.94, EfficientNet: 0.95, EfficientNet+Att: 0.96ACC:ResNet50: 0.78, ResNet50+Att: 0.79, GoogLeNet: 0.81, GoogLeNet+Att: 0.83, EfficientNet: 0.83, EfficientNet+Att: 0.84
[7]	Hematoxylin and Eosin (H&E) stained whole-slide images of lymph node biopsies	BL or DLBCL	Differential diagnosis (BL vs DLBCL)	A total of 10,818 images from BL (n = 34) and DLBCL (n = 36) cases were used to either train or apply different CNNs	Convolutional Neural Network (CNN) in Python3 using Tensorflow	ACC: N3-100:0.876, N4:0.94 AUC:N3-100: DLBCL 0.89 / BL 0.88 N4: DLBCL 0.92 / BL 0.92
[8]	Hematoxylin and eosin (H&E) slides of a lesion area	DLBCL, FL, RLH	Slide level classification into three diagnostic categories	samples of 388 sections composed of 259 DLBCLs, 89 FLs, and 40 RLs, which were nodal and extranodal lesions. All sections were diagnosed at Kurume University from 2010 to 2017.	Deep convolutional neural network (CNN)	ACC:0,970 AUC:DLBCL:0.969 , FL: 1.00 , RLH:0.950
[33]	The H&E-stained slides	+-MYC rearrangement	Slide level classification	internal set of routinely stained H&E glass slides and MYC fluorescence in situ hybridization (FISH) test results of 245 patients that were diagnosed with DLBCL in 11 hospitals in the Netherlands	deep learning neural network (U-Net),Random Forest (RF) classification.	ACC:0.93,Sensitivity:0.90 (internal) 0.95 (external) Specificity: 0.52 (internal) 0.53 (external)
[20]	Slides of lymph nodes stained with H&E	MCL, FL and CLL	Slide level classification into three diagnostic categories	Public dataset from studies conducted by researchers from National Cancer Institute and National Institute on Aging, in the United States. A total of 30 histological slides of lymph nodes stained with H&E	polynomial (PL),random forest (RF), decision tree (DT) and support vector machine developed using Matlab Language with the help of the WEKA platform 3.6.6 for classifying the features. Machine learning algorithms ( J48 module,SMO module)	ACC:up to 1.00 AUC: 0.906 (CLL–MCL),0.891 (CLL–FL),0.859 (FL–MCL)
[9]	Hematoxylin-and-eosin (H&E)-stained tissue slide	DLBCL, AITL, CHL	Slide level classification into three diagnostic categories	From over 80 different institutions database of malignant lymphomas comprises N = 262 clinical cases, which include three subtypes: 67,97, and 98 cases of AITL, DLBCL, and CHL, respectively.	MIL-based CNN with an attention mechanism, utilizing ResNet-50 as a feature extractor	ACC:0.698 (typical), 0.640 (atypical)Macro-F1:0.680 (typical), 0.618 (atypical)
[21]	Digital histopathological images derived from Hematoxylin and Eosin (H&E) stained tissue biopsies	CLL, FL, MCL	Slide level classification into three diagnostic categories	1082 data points where 80% of the data (708) are utilized for training and 20% (374) are used for testing.	Deep learning algorithm with the Bi-LSTM, DBN and RBFN	ACC:0.948Hit Rate:0.900NPV:0.956
[36]	The H&E-stained slides	MALT or Not MALT	Slide level classification	Total of 350 slides, including 106 slides of gastric MALT lymphoma and 244 slides of tumor-free lymphoid tissue.	Multi-model fusion, ResNet50, EfficientNet B0, EfficientNet V2	ACC: ResNet50: 0.9147 EfficientNet B0: 0.7248 EfficientNet V2: 0.8760 fusion: 0.9496
[22]	Whole slide images from bipsies treated with H&E and examined on Liquid Based Cytology (LBC) slides	CLL, FL, MCL	Slide level classification into three diagnostic categories	Dataset from the Kaggle multicancer dataset, comprising 15,000 WSIs, uniformly distributed across three lymphoma categories: FL, CLL and MCL. Each category contributed 5,000 images.	Model HCTN-LC with SqueezeNet and ViT as the architectural backbones	ACC:0.9987,Sensitivity:0.9987,Specifity:0.9993
[10]	H&E-stained formalin-fixed paraffin-embedded tissue sections	DLBCL or not DLBCL	Slide level classification	Pathologic images from three hospitals featuring 1005 images from Hospital A, 3123 from Hospital B, and 402 from Hospital C	GOTDP-MP-CNNs (Globally Optimized Transfer Deep-Learning Platform with Multiple Pretrained CNNs)	ACC:Hospital A:1.0000,Hospital B:0.9971,Hospital C:1.0000
[38]	H&E-stained images	NL, MALT lymphoma, GCB-DLBCL, non-GCB-DLBCL	Patch level classification	Data from 160 patients, comprising 25 normal lymph nodes (NL), 26 MALT lymphoma, 31 GCB, and 78 non-GCB cases purchased from Biomax tissue microarrays (TMAs)	EfficientNet (pretrained on ImageNet), which outperformed 5 other CNNs (AlexNet, VGG16, ResNet18, SqueezeNet, GoogleNet)	ACC: 0.702AUC: 0.870
[23]	H&E histological images	FL, RLH	Patch level classification	Large series of 221 cases, including 177 follicular lymphoma. The series included 1,004,509 follicular lymphoma and 490,506 reactive lymphoid tissue image-patches and 44 reactive lymphoid tissue	Convolutional neural network (CNN) based on ResNet architecture	ACC:0.998, Precision:0.998, Specificity:0.997F1-score: 0.999
[31]	H&E-stained whole slide images	MEITL vs ITCL-NOS	Case-level classification	Total of 40 histopathological whole-slide images (WSIs) from 40 surgically resected PITL cases	Hybrid Model: HTC-RCNN (for nuclear segmentation) + XGBoost (for classification)	Segmentation AP (HTC-RCNN): 0.881 Classification AUC (XGBoost): 0.966 Classification AUC (End-to-end CNN): 0.820
[11]	H&E-stained tissue microarray cores	Agg BCL, DLBCL, FL, CHL, MCL, MZL, NKTCL, TCL	TMA core-level classification	Dataset of 670 cases from Guatemala spanning,8 lymphoma subtypes	LymphoML (LightGBM), SHapley Additive exPlanation (SHAP) analysis	ACC: 0.643
[32]	H&E-stained histopathological images of malignant lymphomas	CLL, FL, MCL	Classification of three malignant lymphoma types	Dataset 1: 15,000 images (5,000 per class). Dataset 2: 374 images (113 CLL, 139 FL, 122 MCL). Train/test split 80/20	System 1: DenseNet-121/ResNet-50 + PCA + SVM. System 2: ResNet-50 + hand-crafted features (GLCM, FCH, DWT, LBP) + FFNN with 756-feature vectors	Dataset 1: ResNet-50 FFNN: ACC: 99.5%, Specificity: 100%, Sensitivity: 99.33%, AUC: 99.86%. Dataset 2: ResNet-50 FFNN: ACC: 100%, Specificity: 100%, Sensitivity: 100%, AUC: 100%
[12]	H&E-stained histopathological images	DLBCL vs Non-DLBCL	Automated diagnosis and classification of DLBCL	1,000 images total (500 DLBCL, 500 Non-DLBCL). Training: 700 images, Validation: 300 images. From lymph nodes only	DDLM-CAM. Two-channel architecture: DenseNet-201 (198 layers) + Attention Map Feature Transformer (AMFT)	ACC: 96%, Recall: 94.67%, Precision: 97.26%, Specificity: 97.33%
[13]	IHC-stained whole-slide images of DLBCL	DLBCL (GCB and ABC subtypes) with PD-L1 expression assessment	Automated quantification of PD-L1 expression and tumor proportion	Primary: 220 patients (88 surgical specimens, 132 fine needle biopsies), 4,101 tissue regions, 146,439 cells annotated. External validation: 61 patients.	ViT-tiny for ROI segmentation, AuxCNN for cell detection, NuClick for cell segmentation. Custom PD-L1 digital quantification rule for DLBCL based on cell morphology and area filtering.	Primary (surgical specimens): ICC Human vs Machine 0.96 (95% CI 0.94-0.97). Fine needle biopsies: ICC 0.96 (95% CI 0.95-0.97). Validation cohort: ICC 0.96 (0.95-0.98) for surgical and 0.98 (95% CI 0.95-0.99) for fine needle biopsies
[14]	IHC c-MYC and BCL2 stained tissue microarray cores	DLBCL with c-MYC and BCL2 positivity assessment	Automated quantification of proportion of c-MYC and BCL2 positive tumor cells from TMAs and WSIs	Training: 378 TMA cores from 173 patients (DLBCL-Morph dataset, Stanford). Validation: 52 WSIs (c-MYC), 56 WSIs (BCL2); 51 patients for double-expressor analysis.	AB-MIL (Attention-based Multiple Instance Learning) with ResNet50 pre-trained on ImageNet for feature extraction.	TMAs - c-MYC: Pearson r 0.843 (95% CI 0.797-0.907), Sensitivity 0.743, Specificity 0.963. BCL2: Pearson r 0.919, Sensitivity 0.938, Specificity 0.951. WSIs - c-MYC: Pearson r 0.883, Sensitivity 0.857-0.706, Specificity 0.991-0.930. BCL2: Pearson r 0.765, Sensitivity 0.856, Specificity 0.690. Double-expressor - WSI: Sensitivity 0.890, Specificity 0.598-1.000
[24]	H&E-stained WSI of lymph nodes	Follicular lymphoma (FL) vs follicular hyperplasia (FH)	Automated differential diagnosis FL vs FH with uncertainty estimation	378 lymph nodes (197 FL, 181 FH); 320,000 patches; train/val/test: 50/25/25%	Bayesian Neural Network (BNN) with patch-based analysis, multiple resolutions, dropout variance for uncertainty, trained at 8 pyramid levels	Patch accuracy: 91%. Slide AUC: 0.92–0.99 (best: lowest resolution). 100% FL detection at 20% false alarm
[37]	H&E-stained frozen whole-slide images of CNS tumors	Primary CNS lymphoma vs glioma (PCNSL) and non-PCNSL	Intraoperative discrimination of PCNSL vs glioma/non-PCNSL	Internal: 432 patients (79 PCNSL, 353 glioma); External 1: 300 (49 PCNSL, 251 glioma); External 2: 386 (22 PCNSL, 364 glioma)	LGNet ensemble (5 ResNet-50), tile-based averaging, patient/slide-level AUC, dropout for uncertainty estimation	External test: AUROC 0.965-0.972 (PCNSL vs glioma), AUROC 0.981-0.993 (PCNSL vs non-PCNSL); Sensitivity up to 95.5%, specificity up to 91.2%
[25]	H&E-stained images, IICBU lymphoma dataset	CLL, FL, MCL	Automated classification of lymphoma subtype	374 images, split into 336 patches per image	Multispace reconstructed images (gradient, GLCM, LBP channels); VGG-16 pretrained, LSTM layer for feature selection, softmax classifier	Patch-level: ACC 98.94%, SEN 96.66–96.85%, SPE 99.12–99.38% (per class)
[26]	H&E-stained WSI	FL, CLL, MCL	Multi-class classification of major lymphoma subtypes	15,000 WSI (5,000 per class, Kaggle), 80/20 train/test split	Hybrid: MobileNet, VGG16, AlexNet (features extracted, fused); handcrafted features (color, wavelet, texture, shape); ACO for selection; XGBoost and DT classifiers	Best: MobileNet-VGG16 + handcrafted + XGBoost: ACC 99.8%, AUC 99.43%, Sens 99.7%, Spec 99.8%, Prec 99.77%
[39]	OCT2 immunostained slides	NLPHL (treatment response prediction)	Prediction of chemotherapy response	53 pediatric patients (14,579 LP cells annotated)	YOLOv4-tiny for cell detection, spatial statistics analysis	Mean AP: 95.24%. LP cell density: p=0.0049; LP cells/cluster: p=0.0012 (good vs poor responders)
[15]	H&E-stained tissue microarrays of lymph node specimens	SLL/CLL and DLBCL	Slide level classification	629 patients (129 SLL/CLL, 119 DLBCL, 381 controls). 84,139 image patches. Train/val/test: 60/20/20	EfficientNet B3 CNN with Adam optimizer	BACC: 95.56% (with quality control). DLBCL: 100% sensitivity/specificity
[40]	FFPE biopsy specimens of Hodgkin lymphoma stained with picrosirius red and MMP9	Hodgkin lymphoma prognosis (qPET-positive vs qPET-negative)	Classification of collagen fiber staining patterns for prognostic prediction	83 cases total (30 for training, 53 for testing). Training set: 7,134 tiles (picrosirius red), 5,788 tiles (MMP9). Test set: 953,068 tiles (picrosirius red), 409,406 tiles (MMP9)	YOLOv4 with transfer learning (pre-trained MS-COCO weights)	AUC: 0.79 (CI: 67.3-91.2). qPET-positive: 18% weakly stained fibers, qPET-negative: 10-14% (p=0.0185). Mean Average Precision: 73.4% (picrosirius red), 87.6% (MMP9)
[16]	H&E-stained biopsy slides from lymph nodes and extranodal tissues	DHL/THL (MYC with BCL2 and/or BCL6 rearrangements) vs non-DH DLBCL	Slide level classification for detection of DHL/THL cases	57 biopsies total (32 training: 5 DHL, 27 non-DH; 25 validation: 10 DHL/THL, 15 non-DH).	pre-trained foundation model on TCGA data and Imagene internal database followed by fine-tuning with Multiple Instance Learning (MIL). Adam optimizer	Sensitivity: 100%, Specificity: 86.7%, Accuracy: 92%, AUC: 0.95
[17]	H&E-stained tissue slides	DLBCL, R-CHOP response prediction	Prediction of immunochemotherapy response	251 WSIs from 216 patients (training and validation sets comprising 80%(200 patients) and a test set consisting of the remaining 51 patients)	Self-supervised learning (DINO ViT-S8) with MIL and TabNet for clinical data integration	AUROC: 0.856 (multi-modal), 0.744 (pathology-only); Sensitivity: 90.2%, Specificity: 70.0%
[18]	H&E-stained lymph node biopsy slides	CLL, aCLL, RT-DLBCL	Classification of CLL disease progression	125 patients (69 CLL, 44 aCLL, 80 RT); 193 slides total, 465 ROIs selected.	Hover-Net (pre-trained on PanNuke) for nuclear segmentation, Random Forest classifier with 4 morphologic and architectural biomarkers	ACC: 82.4% (all 4 biomarkers), AUC: 0.935 (95% CI: 0.797-0.952)
[29]	H&E-stained tissue slides of follicular lymphoma	CB vs non-CB cell classification	Automated centroblast classification for FL grading	500 HPF images from 17 patients; 213 CB and 234 non-CB cell images. Train/test split 80/20.	COB (Classification using Orthogonal Bases) with SVD and CLEM (Classification based on Laplacian Eigenmaps) with nonlinear dimensionality reduction	COB: 99.22 ± 0.75% accuracy, 100% precision/recall. CLEM: 99.07 ± 1.53% accuracy, 100% precision/recall
[28]	H&E-stained lymph node biopsy slides	CLL, aCLL, RT (Richter transformation)	Classification of CLL disease progression	135 patients; 193 slides (69 CLL, 44 aCLL, 80 RT); 465 ROIs total. Train/test split 1:1	Hover-Net nuclear segmentation (pre-trained on PanNuke), Spectral clustering for unsupervised cell phenotyping, XGBoost classifier with 6 cellular features (cell ratios and densities)	Accuracy: 92.5%, AUC: 0.978; Mean accuracy (100 repeated splits): 90.2%, AUC: 0.973
[27]	H&E-stained lymphoma pathological images	CLL, FL, MCL	Classification of three lymphoma types	374 images (CLL, FL, MCL). Train/validation/test split: 6:2:2.	ResNet-50 with residual blocks, batch normalization, ReLU activation, cross-entropy loss.	ACC: 98.63% (ResNet-50). Compared to BP (96%) and GA-BP (97.7%). Paired T-test: p < 0.05 (statistically significant)
[41]	H&E-stained histopathological images	7 lymphoma subtypes (HK-Classical, HK-NLP, NHK-Burkitt, NHK-Follicular, NHK-Mantel, NHK-LargeB-Cell, NHK-TCell)	Classification of rare and aggressive lymphoma subtypes	323 RGB images, ∼46 images per class. 5-fold cross-validation, 80/20 train/test split	ResNet50 with transfer learning, 50 layers, batch normalization, global average pooling, Softmax classifier	Accuracy: 91.6%, Macro-Precision: 92%, Macro-Recall: 91.9%, Macro-F1: 91.9%, Kappa: 0.915
[30]	H&E-stained whole-slide images of PCNSL	PCNSL prognosis prediction	Prognostic prediction of OS and PFS, treatment response, primary resistance	114 patients (68 training, 46 validation); 132 WSIs total. 50 non-overlapping patches per patient.	CellProfiler 4.2.6 for automated feature extraction (802 quantitative features). LASSO-Cox regression model with tenfold cross-validation for Path-score generation. Six ML classifiers (Logistic, KNN, RF, SVM, XGBoost, DT)	Path-score (8-feature): Training cohort OS AUC 0.785 (1-yr), 0.869 (2-yr), 0.973 (3-yr); Validation cohort AUC 0.649 (1-yr), 0.679 (2-yr), 0.733 (3-yr). Nomogram: Training AUC 0.862 (1-yr), 0.932 (2-yr), 0.927 (3-yr); Validation AUC 0.802 (1-yr), 0.768 (2-yr), 0.938 (3-yr)

Table 3. Overview of existing AI models in lymphoma diagnostics.

Ref.	Material	Diagnose	Use of the Model	Dataset	AI architecture / method	Results
[34]	skin biopsies	MF, BIDs	Slide-level classification	private dataset of 924 H&E-stained whole-slide images from skin biopsies: 233 patients with early-stage MF and 353 patients with benign inflammatory dermatoses	clustering-constrained attention multiple instance learning (CLAM), ResNet50, CTransPath, UNI	ACC: CLAM-ResNet50: 0.918, CLAM-CTransPath: 0.921, CLAM-UNI: 0.971
[19]	Bone Marrow	LCT, FL, CLL	Slide-level classification	dataset of 71 patients comprising whole-slide images (WSI) of H&E-stained bone marrow biopsies: 21 patients with follicular lymphoma (FL) and 50 patients with chronic lymphocytic leukemia (CLL)	Convolutional Neural Network (CNN),ResNet-50, MobileNet	AUC: CNN(Ratio Pred)-(FL): 0.881, CNN(End-to-End)-(FL): 0.923, CNN(Ratio Pred)-(CLL): 0.780, CNN(End-to-End)-(CLL): 0.823
[6]	Hematoxylin and Eosin (H&E) stained whole-slide images of lymph nodes	BLN, DLBCL, BL, SLL	Classification into 4 diagnostic categories	Whole-slide images obtained from two public sources: (1) Virtual Pathology at the University of Leeds (355,966 WSIs, 114.92 TB), and (2) Virtual Slide Box from the University of Iowa (>1,000 WSIs). From these, 128 cases (32 per class) were selected.	Convolutional Neural Network (CNN) implemented in Python using TensorFlow and Keras.	ACC:CNN-(Image by Image):0.95, CNN(set-by-set classification using majority voting, 3/5 agreement):1
[35]	Hematoxylin and Eosin (H&E) stained whole-slide images of lymphoid tissue (lymph node and tonsil)	BL	Slide-level classification	160 patients total: 90 BL and 70 control (tonsillectomy or reactive lymphoid samples).	Multiple-Instance Learning (MIL),ResNet50, EfficientNet, and GoogLeNet (Inception-v1)	AUC:ResNet50: 0.90, ResNet50+Att: 0.91, GoogLeNet: 0.93, GoogLeNet+Att: 0.94, EfficientNet: 0.95, EfficientNet+Att: 0.96ACC:ResNet50: 0.78, ResNet50+Att: 0.79, GoogLeNet: 0.81, GoogLeNet+Att: 0.83, EfficientNet: 0.83, EfficientNet+Att: 0.84
[7]	Hematoxylin and Eosin (H&E) stained whole-slide images of lymph node biopsies	BL or DLBCL	Differential diagnosis (BL vs DLBCL)	A total of 10,818 images from BL (n = 34) and DLBCL (n = 36) cases were used to either train or apply different CNNs	Convolutional Neural Network (CNN) in Python3 using Tensorflow	ACC: N3-100:0.876, N4:0.94 AUC:N3-100: DLBCL 0.89 / BL 0.88 N4: DLBCL 0.92 / BL 0.92
[8]	Hematoxylin and eosin (H&E) slides of a lesion area	DLBCL, FL, RLH	Slide level classification into three diagnostic categories	samples of 388 sections composed of 259 DLBCLs, 89 FLs, and 40 RLs, which were nodal and extranodal lesions. All sections were diagnosed at Kurume University from 2010 to 2017.	Deep convolutional neural network (CNN)	ACC:0,970 AUC:DLBCL:0.969 , FL: 1.00 , RLH:0.950
[33]	The H&E-stained slides	+-MYC rearrangement	Slide level classification	internal set of routinely stained H&E glass slides and MYC fluorescence in situ hybridization (FISH) test results of 245 patients that were diagnosed with DLBCL in 11 hospitals in the Netherlands	deep learning neural network (U-Net),Random Forest (RF) classification.	ACC:0.93,Sensitivity:0.90 (internal) 0.95 (external) Specificity: 0.52 (internal) 0.53 (external)
[20]	Slides of lymph nodes stained with H&E	MCL, FL and CLL	Slide level classification into three diagnostic categories	Public dataset from studies conducted by researchers from National Cancer Institute and National Institute on Aging, in the United States. A total of 30 histological slides of lymph nodes stained with H&E	polynomial (PL),random forest (RF), decision tree (DT) and support vector machine developed using Matlab Language with the help of the WEKA platform 3.6.6 for classifying the features. Machine learning algorithms ( J48 module,SMO module)	ACC:up to 1.00 AUC: 0.906 (CLL–MCL),0.891 (CLL–FL),0.859 (FL–MCL)
[9]	Hematoxylin-and-eosin (H&E)-stained tissue slide	DLBCL, AITL, CHL	Slide level classification into three diagnostic categories	From over 80 different institutions database of malignant lymphomas comprises N = 262 clinical cases, which include three subtypes: 67,97, and 98 cases of AITL, DLBCL, and CHL, respectively.	MIL-based CNN with an attention mechanism, utilizing ResNet-50 as a feature extractor	ACC:0.698 (typical), 0.640 (atypical)Macro-F1:0.680 (typical), 0.618 (atypical)
[21]	Digital histopathological images derived from Hematoxylin and Eosin (H&E) stained tissue biopsies	CLL, FL, MCL	Slide level classification into three diagnostic categories	1082 data points where 80% of the data (708) are utilized for training and 20% (374) are used for testing.	Deep learning algorithm with the Bi-LSTM, DBN and RBFN	ACC:0.948Hit Rate:0.900NPV:0.956
[36]	The H&E-stained slides	MALT or Not MALT	Slide level classification	Total of 350 slides, including 106 slides of gastric MALT lymphoma and 244 slides of tumor-free lymphoid tissue.	Multi-model fusion, ResNet50, EfficientNet B0, EfficientNet V2	ACC: ResNet50: 0.9147 EfficientNet B0: 0.7248 EfficientNet V2: 0.8760 fusion: 0.9496
[22]	Whole slide images from bipsies treated with H&E and examined on Liquid Based Cytology (LBC) slides	CLL, FL, MCL	Slide level classification into three diagnostic categories	Dataset from the Kaggle multicancer dataset, comprising 15,000 WSIs, uniformly distributed across three lymphoma categories: FL, CLL and MCL. Each category contributed 5,000 images.	Model HCTN-LC with SqueezeNet and ViT as the architectural backbones	ACC:0.9987,Sensitivity:0.9987,Specifity:0.9993
[10]	H&E-stained formalin-fixed paraffin-embedded tissue sections	DLBCL or not DLBCL	Slide level classification	Pathologic images from three hospitals featuring 1005 images from Hospital A, 3123 from Hospital B, and 402 from Hospital C	GOTDP-MP-CNNs (Globally Optimized Transfer Deep-Learning Platform with Multiple Pretrained CNNs)	ACC:Hospital A:1.0000,Hospital B:0.9971,Hospital C:1.0000
[38]	H&E-stained images	NL, MALT lymphoma, GCB-DLBCL, non-GCB-DLBCL	Patch level classification	Data from 160 patients, comprising 25 normal lymph nodes (NL), 26 MALT lymphoma, 31 GCB, and 78 non-GCB cases purchased from Biomax tissue microarrays (TMAs)	EfficientNet (pretrained on ImageNet), which outperformed 5 other CNNs (AlexNet, VGG16, ResNet18, SqueezeNet, GoogleNet)	ACC: 0.702AUC: 0.870
[23]	H&E histological images	FL, RLH	Patch level classification	Large series of 221 cases, including 177 follicular lymphoma. The series included 1,004,509 follicular lymphoma and 490,506 reactive lymphoid tissue image-patches and 44 reactive lymphoid tissue	Convolutional neural network (CNN) based on ResNet architecture	ACC:0.998, Precision:0.998, Specificity:0.997F1-score: 0.999
[31]	H&E-stained whole slide images	MEITL vs ITCL-NOS	Case-level classification	Total of 40 histopathological whole-slide images (WSIs) from 40 surgically resected PITL cases	Hybrid Model: HTC-RCNN (for nuclear segmentation) + XGBoost (for classification)	Segmentation AP (HTC-RCNN): 0.881 Classification AUC (XGBoost): 0.966 Classification AUC (End-to-end CNN): 0.820
[11]	H&E-stained tissue microarray cores	Agg BCL, DLBCL, FL, CHL, MCL, MZL, NKTCL, TCL	TMA core-level classification	Dataset of 670 cases from Guatemala spanning,8 lymphoma subtypes	LymphoML (LightGBM), SHapley Additive exPlanation (SHAP) analysis	ACC: 0.643
[32]	H&E-stained histopathological images of malignant lymphomas	CLL, FL, MCL	Classification of three malignant lymphoma types	Dataset 1: 15,000 images (5,000 per class). Dataset 2: 374 images (113 CLL, 139 FL, 122 MCL). Train/test split 80/20	System 1: DenseNet-121/ResNet-50 + PCA + SVM. System 2: ResNet-50 + hand-crafted features (GLCM, FCH, DWT, LBP) + FFNN with 756-feature vectors	Dataset 1: ResNet-50 FFNN: ACC: 99.5%, Specificity: 100%, Sensitivity: 99.33%, AUC: 99.86%. Dataset 2: ResNet-50 FFNN: ACC: 100%, Specificity: 100%, Sensitivity: 100%, AUC: 100%
[12]	H&E-stained histopathological images	DLBCL vs Non-DLBCL	Automated diagnosis and classification of DLBCL	1,000 images total (500 DLBCL, 500 Non-DLBCL). Training: 700 images, Validation: 300 images. From lymph nodes only	DDLM-CAM. Two-channel architecture: DenseNet-201 (198 layers) + Attention Map Feature Transformer (AMFT)	ACC: 96%, Recall: 94.67%, Precision: 97.26%, Specificity: 97.33%
[13]	IHC-stained whole-slide images of DLBCL	DLBCL (GCB and ABC subtypes) with PD-L1 expression assessment	Automated quantification of PD-L1 expression and tumor proportion	Primary: 220 patients (88 surgical specimens, 132 fine needle biopsies), 4,101 tissue regions, 146,439 cells annotated. External validation: 61 patients.	ViT-tiny for ROI segmentation, AuxCNN for cell detection, NuClick for cell segmentation. Custom PD-L1 digital quantification rule for DLBCL based on cell morphology and area filtering.	Primary (surgical specimens): ICC Human vs Machine 0.96 (95% CI 0.94-0.97). Fine needle biopsies: ICC 0.96 (95% CI 0.95-0.97). Validation cohort: ICC 0.96 (0.95-0.98) for surgical and 0.98 (95% CI 0.95-0.99) for fine needle biopsies
[14]	IHC c-MYC and BCL2 stained tissue microarray cores	DLBCL with c-MYC and BCL2 positivity assessment	Automated quantification of proportion of c-MYC and BCL2 positive tumor cells from TMAs and WSIs	Training: 378 TMA cores from 173 patients (DLBCL-Morph dataset, Stanford). Validation: 52 WSIs (c-MYC), 56 WSIs (BCL2); 51 patients for double-expressor analysis.	AB-MIL (Attention-based Multiple Instance Learning) with ResNet50 pre-trained on ImageNet for feature extraction.	TMAs - c-MYC: Pearson r 0.843 (95% CI 0.797-0.907), Sensitivity 0.743, Specificity 0.963. BCL2: Pearson r 0.919, Sensitivity 0.938, Specificity 0.951. WSIs - c-MYC: Pearson r 0.883, Sensitivity 0.857-0.706, Specificity 0.991-0.930. BCL2: Pearson r 0.765, Sensitivity 0.856, Specificity 0.690. Double-expressor - WSI: Sensitivity 0.890, Specificity 0.598-1.000
[24]	H&E-stained WSI of lymph nodes	Follicular lymphoma (FL) vs follicular hyperplasia (FH)	Automated differential diagnosis FL vs FH with uncertainty estimation	378 lymph nodes (197 FL, 181 FH); 320,000 patches; train/val/test: 50/25/25%	Bayesian Neural Network (BNN) with patch-based analysis, multiple resolutions, dropout variance for uncertainty, trained at 8 pyramid levels	Patch accuracy: 91%. Slide AUC: 0.92–0.99 (best: lowest resolution). 100% FL detection at 20% false alarm
[37]	H&E-stained frozen whole-slide images of CNS tumors	Primary CNS lymphoma vs glioma (PCNSL) and non-PCNSL	Intraoperative discrimination of PCNSL vs glioma/non-PCNSL	Internal: 432 patients (79 PCNSL, 353 glioma); External 1: 300 (49 PCNSL, 251 glioma); External 2: 386 (22 PCNSL, 364 glioma)	LGNet ensemble (5 ResNet-50), tile-based averaging, patient/slide-level AUC, dropout for uncertainty estimation	External test: AUROC 0.965-0.972 (PCNSL vs glioma), AUROC 0.981-0.993 (PCNSL vs non-PCNSL); Sensitivity up to 95.5%, specificity up to 91.2%
[25]	H&E-stained images, IICBU lymphoma dataset	CLL, FL, MCL	Automated classification of lymphoma subtype	374 images, split into 336 patches per image	Multispace reconstructed images (gradient, GLCM, LBP channels); VGG-16 pretrained, LSTM layer for feature selection, softmax classifier	Patch-level: ACC 98.94%, SEN 96.66–96.85%, SPE 99.12–99.38% (per class)
[26]	H&E-stained WSI	FL, CLL, MCL	Multi-class classification of major lymphoma subtypes	15,000 WSI (5,000 per class, Kaggle), 80/20 train/test split	Hybrid: MobileNet, VGG16, AlexNet (features extracted, fused); handcrafted features (color, wavelet, texture, shape); ACO for selection; XGBoost and DT classifiers	Best: MobileNet-VGG16 + handcrafted + XGBoost: ACC 99.8%, AUC 99.43%, Sens 99.7%, Spec 99.8%, Prec 99.77%
[39]	OCT2 immunostained slides	NLPHL (treatment response prediction)	Prediction of chemotherapy response	53 pediatric patients (14,579 LP cells annotated)	YOLOv4-tiny for cell detection, spatial statistics analysis	Mean AP: 95.24%. LP cell density: p=0.0049; LP cells/cluster: p=0.0012 (good vs poor responders)
[15]	H&E-stained tissue microarrays of lymph node specimens	SLL/CLL and DLBCL	Slide level classification	629 patients (129 SLL/CLL, 119 DLBCL, 381 controls). 84,139 image patches. Train/val/test: 60/20/20	EfficientNet B3 CNN with Adam optimizer	BACC: 95.56% (with quality control). DLBCL: 100% sensitivity/specificity
[40]	FFPE biopsy specimens of Hodgkin lymphoma stained with picrosirius red and MMP9	Hodgkin lymphoma prognosis (qPET-positive vs qPET-negative)	Classification of collagen fiber staining patterns for prognostic prediction	83 cases total (30 for training, 53 for testing). Training set: 7,134 tiles (picrosirius red), 5,788 tiles (MMP9). Test set: 953,068 tiles (picrosirius red), 409,406 tiles (MMP9)	YOLOv4 with transfer learning (pre-trained MS-COCO weights)	AUC: 0.79 (CI: 67.3-91.2). qPET-positive: 18% weakly stained fibers, qPET-negative: 10-14% (p=0.0185). Mean Average Precision: 73.4% (picrosirius red), 87.6% (MMP9)
[16]	H&E-stained biopsy slides from lymph nodes and extranodal tissues	DHL/THL (MYC with BCL2 and/or BCL6 rearrangements) vs non-DH DLBCL	Slide level classification for detection of DHL/THL cases	57 biopsies total (32 training: 5 DHL, 27 non-DH; 25 validation: 10 DHL/THL, 15 non-DH).	pre-trained foundation model on TCGA data and Imagene internal database followed by fine-tuning with Multiple Instance Learning (MIL). Adam optimizer	Sensitivity: 100%, Specificity: 86.7%, Accuracy: 92%, AUC: 0.95
[17]	H&E-stained tissue slides	DLBCL, R-CHOP response prediction	Prediction of immunochemotherapy response	251 WSIs from 216 patients (training and validation sets comprising 80%(200 patients) and a test set consisting of the remaining 51 patients)	Self-supervised learning (DINO ViT-S8) with MIL and TabNet for clinical data integration	AUROC: 0.856 (multi-modal), 0.744 (pathology-only); Sensitivity: 90.2%, Specificity: 70.0%
[18]	H&E-stained lymph node biopsy slides	CLL, aCLL, RT-DLBCL	Classification of CLL disease progression	125 patients (69 CLL, 44 aCLL, 80 RT); 193 slides total, 465 ROIs selected.	Hover-Net (pre-trained on PanNuke) for nuclear segmentation, Random Forest classifier with 4 morphologic and architectural biomarkers	ACC: 82.4% (all 4 biomarkers), AUC: 0.935 (95% CI: 0.797-0.952)
[29]	H&E-stained tissue slides of follicular lymphoma	CB vs non-CB cell classification	Automated centroblast classification for FL grading	500 HPF images from 17 patients; 213 CB and 234 non-CB cell images. Train/test split 80/20.	COB (Classification using Orthogonal Bases) with SVD and CLEM (Classification based on Laplacian Eigenmaps) with nonlinear dimensionality reduction	COB: 99.22 ± 0.75% accuracy, 100% precision/recall. CLEM: 99.07 ± 1.53% accuracy, 100% precision/recall
[28]	H&E-stained lymph node biopsy slides	CLL, aCLL, RT (Richter transformation)	Classification of CLL disease progression	135 patients; 193 slides (69 CLL, 44 aCLL, 80 RT); 465 ROIs total. Train/test split 1:1	Hover-Net nuclear segmentation (pre-trained on PanNuke), Spectral clustering for unsupervised cell phenotyping, XGBoost classifier with 6 cellular features (cell ratios and densities)	Accuracy: 92.5%, AUC: 0.978; Mean accuracy (100 repeated splits): 90.2%, AUC: 0.973
[27]	H&E-stained lymphoma pathological images	CLL, FL, MCL	Classification of three lymphoma types	374 images (CLL, FL, MCL). Train/validation/test split: 6:2:2.	ResNet-50 with residual blocks, batch normalization, ReLU activation, cross-entropy loss.	ACC: 98.63% (ResNet-50). Compared to BP (96%) and GA-BP (97.7%). Paired T-test: p < 0.05 (statistically significant)
[41]	H&E-stained histopathological images	7 lymphoma subtypes (HK-Classical, HK-NLP, NHK-Burkitt, NHK-Follicular, NHK-Mantel, NHK-LargeB-Cell, NHK-TCell)	Classification of rare and aggressive lymphoma subtypes	323 RGB images, ∼46 images per class. 5-fold cross-validation, 80/20 train/test split	ResNet50 with transfer learning, 50 layers, batch normalization, global average pooling, Softmax classifier	Accuracy: 91.6%, Macro-Precision: 92%, Macro-Recall: 91.9%, Macro-F1: 91.9%, Kappa: 0.915
[30]	H&E-stained whole-slide images of PCNSL	PCNSL prognosis prediction	Prognostic prediction of OS and PFS, treatment response, primary resistance	114 patients (68 training, 46 validation); 132 WSIs total. 50 non-overlapping patches per patient.	CellProfiler 4.2.6 for automated feature extraction (802 quantitative features). LASSO-Cox regression model with tenfold cross-validation for Path-score generation. Six ML classifiers (Logistic, KNN, RF, SVM, XGBoost, DT)	Path-score (8-feature): Training cohort OS AUC 0.785 (1-yr), 0.869 (2-yr), 0.973 (3-yr); Validation cohort AUC 0.649 (1-yr), 0.679 (2-yr), 0.733 (3-yr). Nomogram: Training AUC 0.862 (1-yr), 0.932 (2-yr), 0.927 (3-yr); Validation AUC 0.802 (1-yr), 0.768 (2-yr), 0.938 (3-yr)

4. Discussion

This scoping review presents a systematic effort to map the rapidly evolving landscape of artificial intelligence applications in lymphoma histopathology diagnostics, spanning tasks from classification of rare subtypes and prediction of genetic rearrangements to objective prognostic modelling and quantitative assessment of immunohistochemical biomarkers [13,14,17,30,33]. The main emphasis of the discussion is placed on a critical appraisal of the nature and quality of existing models and on identifying key translational gaps that still hinder widespread deployment of these advanced systems in routine hematopathology practice [42].

4.1. Characterization of AI Task Types in Hematopathology

Analysis of the assembled evidence supports three major pillars underpinning contemporary AI tasks in hematopathology: morphological classification, advanced discrimination of overlapping entities (difficult differentials), quantitative–predictive tasks. A foundational class of tasks is lymphoma subtype classification on routinely stained H&E whole-slide images [15,26,32]. These models have progressed from simple binary classifiers to multiclass systems capable of recognizing key entities such as diffuse large B-cell lymphoma (DLBCL), follicular lymphoma (FL), mantle cell lymphoma (MCL), and chronic lymphocytic leukemia/small lymphocytic lymphoma (CLL/SLL) [15,26,32].

Particular clinical value is attributed to models designed for ”difficult differentials”, where morphology is highly non-specific. Examples include discrimination of Burkitt lymphoma (BL) from DLBCL [7], therapeutically critical, as well as FL versus benign reactive follicular hyperplasia (FH) [23,24]. At the same time, concerns remain regarding generalizability across sites, as performance can be highly sensitive to inter-institutional technical heterogeneity (e.g., staining and scanning variation), with substantial degradation outside the training domain [24].

A separate, rapidly developing subgroup comprises disease-progression assessment models, particularly in indolent lymphomas such as CLL, where AI may help objectify recognition of large-cell (Richter) transformation and accelerated-phase CLL (aCLL), which are known to be subject to marked interobserver variability in routine microscopy [19,28].

4.2. Network Architectures and Learning Paradigms

Technological trajectory The technological evolution of AI models in hematopathology traces a clear trajectory from relatively simple morphometric classifiers toward advanced systems that integrate multiple data modalities.

Early and still common approaches are based on classical AI models such as Decision Tree (DT) [20,31], Random Forest (RF) [18,20] and Support Vector Machine (SVM) [32]. However, the most common approaches are based on classical convolutional neural network (CNN) architectures, such as ResNet, EfficientNet, GoogLeNet, MobileNet, and DenseNet [27]. Nevertheless, classical CNNs frequently face the ”black-box” limitation and a constrained ability to model global architectural dependencies across the entire lymph node [22].

The specific characteristics of lymphoma histopathology, such as often diffuse infiltration, subtle microenvironmental changes, and high tissue heterogeneity, have driven a shift toward the Multiple Instance Learning (MIL) paradigm [14]. MIL frameworks, including the Clustering-constrained Attention Multiple Instance Learning (CLAM) algorithm, are well suited to WSI analysis because they aggregate information across thousands of patches using only slide-level labels, avoiding the need for dense pixel-level manual annotation.

Attention mechanisms enable the network to focus on diagnostically salient regions while reducing the importance of artefacts and irrelevant background, which is critical when detecting rare cellular populations or events such as Richter transformation [28]. A major step towards global histopathology understanding and an extension of attention mechanisms in image analysis is the introduction of Vision Transformers (ViT) and CNN–ViT hybrid models [22]. Although this is a relatively new approach, models such as HCTN-LC [22] are already emerging, combining a lightweight SqueezeNet with a ViT branch, aim to capture spatial dependencies that are difficult to represent with classical convolutions alone. On the other hand, using smaller variants such as ViT-Tiny has also been reported as effective for precise ROI segmentation in immunohistochemistry, supporting more objective assessment of markers such as PD-L1 [13].

One of the most promising trends beyond the constraints of small annotated datasets is the rise of self-supervised learning and foundation models [35]. Using paradigms such as DINO with ViT backbones (e.g., DINO-ViT / DINO ViT-S) trained on large collections of unlabeled slides enables extraction of more generalizable morphological representations. These representations can serve as a robust basis for transfer learning across institutions, mitigating the impact of technical heterogeneity [16].

4.3. Training Data and Their Sources

A key determinant of the translational success of AI models in hematopathology is the quality, scale, and provenance of the training data. An analysis of the reviewed literature reveals a pronounced dichotomy between small, meticulously curated clinical cohorts and large-scale, yet often more weakly characterized, public datasets.

On the one hand, there are very small, single-centre series using as few as 30 – 40 WSIs or only several dozen clinical cases, which is typical for rare entities or proof-of-concept studies [20,31]. On the other hand, systems trained on resources such as Kaggle may operate on datasets on the order of 15,000 images, while the most advanced analyses of lymph node architecture process even millions of patches [22,23].

The type of biological material analyzed is similarly heterogeneous, forcing models to contend with distinct artefact profiles:

1.

Lymph node biopsies and TMA: These remain the standard, often in the form of tissue microarrays (TMAs), which reduce search variance but constrain architectural context [43].

2.

Bone marrow: Used mainly for assessing large-cell transformation, where single-cell segmentation in densely cellular substrates is critical [19].

3.

Skin biopsies: Models dedicated to cutaneous T-cell lymphomas (MF) must distinguish neoplastic infiltrates from benign dermatoses [34].

4.

Frozen sections: Intraoperative assessment (e.g., in PCNSL) is affected by specific freezing artefacts, which models such as LGNet can effectively disregard [37].

5.

LBC cytology: Liquid-based cytology enables analysis of isolated cells, which simplifies segmentation but sacrifices tissue-architecture information [26].

Public datasets, such as NCI/NIA, Kaggle, Leeds/Iowa virtual pathology resources, or commercial TMA products (e.g., BioMax), have played a fundamental role in enabling study reproducibility [38]. Crucially, these resources allow objective comparisons of different network architectures, as illustrated by studies benchmarking EfficientNet against classical CNN baselines [44]

Overreliance on private, institution-specific datasets introduces the risk of ”institutional overfitting”. Models trained on such datasets may achieve excellent internal performance that deteriorates sharply on external data [24]. Key technical confounders include scanner characteristics, such as sensor and compression differences, and variability in staining protocols, particularly in hematoxylin and eosin intensity [10]. In addition, selection bias may occur, resulting in private datasets overrepresent “textbook” while underrepresenting borderline cases and artefacts, thereby artificially simplifying the classification problem [16].

4.4. Interpretation of Performance Metrics

Interpreting AI model performance metrics in hematopathology requires moving beyond a superficial reading of high numerical values, which often conceal material methodological risks. In the reviewed literature, it is striking how frequently studies report near-perfect results, such as ACC or AUC close to 1.0. Particularly for tasks with relatively low morphological complexity or for carefully curated, small datasets [22,32]. Such results may be affected by ”optimism bias” driven by limited sample representativeness. In a systematic review covering several dozen models, this was identified as a high risk of bias in participant selection and statistical analysis domains [42].

Although metrics such as AUC and ACC are standard in machine learning, lymphoma diagnostics increasingly benefits from measures that are more natural to pathology practice and better reflect tool utility in precision medicine:

1.

Intraclass correlation coefficient (ICC): For PD-L1 expression assessment in DLBCL, ACC alone is insufficient due to tissue heterogeneity; therefore, reporting ICC is essential, with AI algorithms achieving ICC = 0.96, exceeding inter-pathologist agreement (0.94) [13].

2.

Pearson correlation: In quantitative scoring of c-MYC and BCL2, AI outputs achieve correlations with the gold standard of 0.843 and 0.919, respectively, enabling more objective identification of the double-expressor phenotype than is achievable with binary classifiers [14].

The key test of model robustness is not performance on an internal test set, but external validation, which exposes AI sensitivity to technical variation. Li et al. show that a model achieving 100% accuracy in one institution, when transferred to another site, loses nearly 10% precision (dropping to 90.5%) solely due to a lack of harmonization in image geometry and acquisition parameters [10]. Even more sceptical conclusions emerge from studies of Bayesian Neural Networks (BNN): despite ideal internal validation (AUC = 1.0), external performance dropped markedly (AUC 0.63–0.69), which Syrykh et al. attribute to high sensitivity to pre-processing protocols [24].

Both clinical experience and the literature point to a fundamental lack of standardization that prevents straightforward, head-to-head ranking of models. Studies vary in:

1.

Decision thresholds: For example, different cut-offs for Ki67 (40% to 8%) or PD-L1 (1%, 5%, 50%) can drastically alter sensitivity and specificity [13,38].

2.

ROI selection strategies: Some models operate on manually selected “clean” regions of interest (ROIs), which can artificially inflate metrics, whereas others analyze entire WSIs and thus confront real-world diagnostic noise [13,28].

3.

Class definitions: For instance, the DLBCL category in one study may include only the NOS subtype, while another also includes aggressive variants, which changes the fundamental level of task difficulty [7,17].

Accordingly, inferences at the discussion level must be formulated with caution: high ACC/AUC should be treated more as evidence that biologically meaningful signal is present in the image than as a certificate of clinical readiness of the tool [42].

4.5. Models for IHC and Prognostic Markers

AI systems appear most readily implementable in lymphoma histopathology when used to automate IHC biomarker assessment, including PD-L1, c-MYC, and BCL2. In routine practice, scoring of these markers is subjective and shows interobserver variability that can affect treatment decisions [13,14]. AI can reduce this variability by providing quantitative, standardized measurements that support precision medicine [13].

Yan et al. demonstrated this approach for PD-L1 in DLBCL using more than 146,000 individual cells from 5,101 tissue regions, enabling rule-based filtering of non-tumour cells such as macrophages, an important issue when neoplastic and reactive populations may share antigen expression [13]. Importantly, the models were trained and evaluated using clinically accepted decision thresholds (e.g., 5% or 50% TPS), which makes outputs directly interpretable for clinicians [13].

Evidence also suggests robustness across sample preparations. AI achieved higher agreement with pathologists on FNA biopsies than on large surgical specimens, plausibly due to lower image noise in smaller samples [13]. Tavolara et al. further showed that models trained on standardized standardized tissue microarrays (TMA) can be transferred to WSI analysis for c-MYC/BCL2, addressing staining and architectural heterogeneity and outperforming pathologists in g progression-free 486 survival (PFS) risk stratification among double-expressor patients [14].

Because these systems output clinically actionable numeric scores aligned with existing decision pathways (e.g., WHO-oriented workflows), they are strong candidates for rapid translation, for example as an automated “second read” integrated into routine IHC processing [13,14].

4.6. The Gap Between Research Findings and Approved Solutions

Despite the breadth of published research, a clear gap remains between scientific output and real-world clinical implementation, as reflected in regulatory registers. At present, the list of FDA-cleared AI solutions in pathology includes only six entries, none of which directly address lymphoma histopathology[1]. This disparity can be explained by several translational barriers.

A central limitation is the lack of prospective clinical evidence. Nearly all reviewed AI models were evaluated retrospectively, which does not capture the variability of routine workflows [1,42]. Retrospective validation fails to account for real-world factors such as differences in specimen preparation or scanning artefacts; therefore, prospective studies with hard clinical endpoints remain an urgent requirement for regulatory acceptance [1,42].

Limited generalizability is further driven by pipeline variation and insufficient standardization across the diagnostic process, from scanner parameters and staining protocols to harmonized quality metrics [38]. Technical variance introduced during slide preparation can markedly reduce inter-institutional performance, leading to site-tailored rather than universally deployable models [10]. These models are highly sensitive to pre-processing, where even minor deviations in hematoxylin intensity may trigger misclassification [24].

Another barrier is the inherent subjectivity of pathologists themselves, which propagates into training-label quality. In challenging lymphoma differential diagnosis settings, interobserver variability may range from 15% to as high as 40% [1]. For quantification of markers such as c-MYC or BCL2, the subjective nature of scoring introduces label noise, complicating the definition of what constitutes a “correct” classification in borderline cases [14,29]. Without harmonized endpoint definitions, rigorous comparison of different AI systems remains difficult [42].

Finally, technological maturity should be separated from clinical-regulatory maturity: high laboratory performance (e.g., AUC > 0.95) does not imply readiness for deployment [42]. Moving toward consolidation requires transparent reporting of participant recruitment and adherence to guidelines for trustworthy medical AI [42]. Accordingly, the field appears to be transitioning from proof-of-concept toward clinical validation, which is necessary for AI to become a routine tool supporting lymphoma diagnostics [13].

4.7. Specific Limitations of AI Models in Lymphoma Histopathology

Despite promising metrics reported under controlled experimental conditions, the assembled evidence highlights limitations that still hinder translation into routine hematopathology practice. These constraints mainly arise from the morphological complexity and incomplete diagnostic context with limited interpretability, compounded by persistent risks of bias and poor generalization.

4.7.1. Morphological Complexity

A fundamental barrier is the extreme morphological complexity of lymphomas, with overlapping cytologic features between benign and malignant entities, which is most apparent in difficult differentials such as FL versus FH [24]. While models perform well in “typical” cases, performance drops in the diagnostic grey zone, where subtle architectural changes or technical artefacts (e.g., crush artefact, fixation-related errors) lead to misclassification [24,31]. In entities such as monomorphic epitheliotropic intestinal T-cell lymphoma (MEITL), discordance between immunophenotype and atypical nuclear morphology further increases the risk of overinterpretation by current systems [31].

4.7.2. Incomplete Capture of Diagnostic Context

Another key limitation is that many models isolate morphology from the broader clinicopathological context: they rely on a single stain (H&E or a specific IHC marker), although final diagnosis integrates immunophenotypic panels, molecular testing (e.g. FISH), and clinical data [14]. Although integrating histology with clinical variables (e.g. TabNet) can improve prediction of immunochemotherapy response, multimodal approaches are still not the default [17]. From the clinician’s perspective, limited transparency of the decision process remains a major barrier to trust [28]; only a minority of studies apply XAI (e.g. SHAP, Grad-CAM, occlusion sensitivity) in a way that allows verification that models attend to biologically meaningful features rather than noise [11,23]. This need is particularly acute for tasks such as predicting CLL transformation to Richter transformation (RT) or identifying DHL/THL, where nuclear morphology provides key cues that AI should be able to justify [16,28].

4.7.3. Bias Risk and Limited Generalization

Finally, there remains a material risk of data-selection bias and limited generalization: training data often come from specialized centres [42], may underrepresent certain regions or ethnic groups [1,11,35], and may not match real-world specimen types such as FNA or trephine biopsies where artefacts and noise are more frequent [13,35]. Accordingly, rigorous reporting of participant recruitment and statistical transparency remains necessary to close the gap between technological performance and clinical trustworthiness [42].

4.8. Limitations of the Review

In line with the scoping review protocol, this work did not conduct a rigorous, study-by-study methodological quality appraisal with respect to risk of bias. This limits the ability to make evidence-strength claims beyond mapping the landscape, especially given that reporting in medical AI is often insufficient for confident critical appraisal [42].

A further limitation is reliance on performance metrics (ACC, AUC) as reported directly by study authors [42]. Because primary studies frequently provide incomplete methodological detail (e.g., recruitment procedures or whether predictor assessment was blinded to outcomes), synthesis may be affected by uncertainty about how reported results were obtained and how comparable they are across studies [42].

Methodological heterogeneity across the reviewed studies prevents construction of a reliable model ranking. Differences in diagnostic class definitions, validation protocols, and frequent lack of publicly available source code limit reproducibility and impede direct comparisons [23,42]. In addition, confounding from scanner- and slide-preparation-related technical variance, often not adequately mitigated via normalization, restricts how confidently conclusions can be generalized across clinical environments [10,24,34].

Finally, by intentionally restricting the scope of the review to H&E and IHC histopathology images, the full multimodal diagnostic pathway for lymphoma is not assessed. This limits the ability to comment on integrated decision-support systems combining morphology with immunophenotyping, flow cytometry, genomics, radiomics, or other -omics data [17,30,42].

4.9. Perspectives and Actionable Future Research Directions

The current evidence base suggests that AI in hematopathology remains saturated with proof-of-concept–level studies, and future work should therefore prioritize rigorous clinical and operational validation. Future progress can be framed around four directions, with some themes (e.g., prospective validation and generalizability) already emphasized earlier in this discussion.

4.9.1. Shift Toward Prospective, Multi-Centre Studies

A core challenge remains the “generalizability gap” driven by technical inter-laboratory variance. Li et al. clearly demonstrated that a model achieving 100% accuracy in one hospital can lose 10% precision in another if variability arising from staining protocols and scanner parameters is not addressed[10]. Accordingly, prospective studies embedded in real-world pathology workflows are urgently needed to quantify the impact of real-time artefacts (e.g., crush artefact, air bubbles) on prediction stability [35,42]. Only such validation can determine whether AI models genuinely shorten time-to-diagnosis and reduce the need for costly external consultations .

4.9.2. Public Benchmarks and Standardization

Model-to-model comparison is currently hampered by the scarcity of publicly available, high-quality datasets (with source code accessible in only ~22% of publications) [42]. There is a clear need for large public benchmarks targeting clinically contentious tasks:

1.

FL vs FH: This task is associated with ~25% interobserver variability, making it an ideal testing ground for AI [29].

2.

BL vs DLBCL: Models must demonstrate robustness to rare morphological variants [7].

3.

DHL/THL detection: AI as a triage tool prior to costly FISH testing could yield measurable economic benefits [16].

4.9.3. Foundation Models and Multimodal Integration

Emerging paradigms such as foundation models and self-supervised learning may reduce dependence on small labeled datasets [17,34]. However, the central goal for precision hematopathology remains multimodal integration: Lee et al. showed that combining H&E-derived features with clinical and molecular variables (e.g., via TabNet) can increase AUC from 0.744 to 0.856 for predicting R-CHOP response [17]. Future systems should therefore focus on synthesizing morphology, immunophenotype, and genetics to better support therapy personalization.

4.9.4. Redefining Success: From Metrics to Utility

The field should shift from reporting benchmark superiority to demonstrating reliable support for clinical decisions in a defined workflow [42]. The absence of FDA-authorized solutions directly for lymphoma reflects, in part, limited evidence for clinically or economically meaningful benefit[20].

Accordingly, success should be defined by utility endpoints, such as improved outcomes (PFS/OS) or safe de-escalation of parts of an antibody panel without loss of diagnostic quality [14,45], ideally supported by interpretable outputs that help build clinical trust.