Artificial Intelligence for Low-Dose CT Lung Cancer Screening: A Systematic Review of Deep Learning, Input Dimensionality, Workflow Integration, and Foundation Models

1. Introduction

Lung cancer remains the leading cause of cancer-related death globally, and the overall 5-year survival rate is only around 20–25%, largely because most cases are diagnosed at an advanced stage [1]. The National Lung Screening Trial (NLST) established that annual screening of high-risk individuals with low-dose computed tomography (LDCT) reduces lung cancer mortality by about 20% relative to chest radiography; the trial randomized 53,454 participants, of whom 26,722 were assigned to the LDCT arm—an enrollment scale that remains the largest screening CT cohort and that shapes much of the data-availability discussion in this review [2]. The Dutch–Belgian NELSON trial subsequently reported, at ten years of follow-up, a 24% reduction in lung-cancer mortality among screened high-risk men (rate ratio 0.76), with subgroup data suggesting an even larger but less precisely estimated benefit in women (rate ratio 0.67) [3]; the trial’s earlier reports document its volume-based nodule-management protocol [4]. These landmark trials also exposed two persistent problems: a high false-positive rate (over 90% of nodules flagged in NLST were not cancer), and the workload that screening at population scale imposes on a finite radiology workforce.

In current practice, guidelines recommend annual LDCT for high-risk adults, though their eligibility criteria differ. The U.S. Preventive Services Task Force (2021) recommends screening adults aged 50–80 years with a ≥20 pack-year smoking history who currently smoke or quit within the past 15 years [5], whereas the American Cancer Society’s 2023 update broadened eligibility to adults aged 50–80 years who currently or formerly smoked and have a ≥20 pack-year history, removing the years-since-quitting requirement [6]. Each LDCT is read slice-by-slice—often hundreds of axial images (0.6–2.5 mm thick) per case—and findings are categorized using systems such as Lung-RADS. Benign processes (scars, lymph nodes, infections) can mimic cancer, and in NLST roughly a quarter of participants had a positive finding requiring follow-up, of which the large majority were not cancer. The sheer number of images requiring scrutiny, combined with the small fraction that represent early cancer, is itself a workload problem and a strong motivation for AI assistance.

Since 2018, convolutional neural networks (CNNs) have become central to medical image analysis, and groups have reported expert-level nodule detection and classification [7,8,9,10]. More recently, transformer architectures and large self-supervised “foundation models” have been explored for lung imaging [11,12,13]. Deep learning is now pervasive across oncologic imaging more broadly—spanning digital pathology [14], transfer-learning–based classification in other tumor types [15,16], cloud-deployed detection pipelines [17], and fusion detection–segmentation systems [18]—and prior reviews and representative lung-cancer studies summarize its growing role in radiology, lung cancer, lung digital pathology, and medical image segmentation [7,9,14,19,21,22,33]. Despite this progress, a practical gap remains: most work relies on supervised learning from relatively small labeled datasets, and there is little consensus on how to choose input dimensionality (2D, 2.5D, or 3D) and model family (CNN vs. transformer) for the data-limited, class-imbalanced screening setting.

Several reviews have surveyed AI in lung cancer broadly [7,19]. Unlike prior reviews that summarize applications across detection, diagnosis, and prognosis, the present review focuses specifically on the screening setting—LDCT in asymptomatic high-risk adults—and organizes the evidence around four design decisions that determine whether AI improves the screening benefit–harm balance: input dimensionality (2D/2.5D/3D), model family (CNN vs. transformer), foundation-model readiness, and reader-facing workflow integration. To our knowledge, no prior review ties these choices to the small-data, class-imbalanced statistics specific to screening or to concrete reductions in false positives, missed cancers, and radiologist reading burden.

This systematic review evaluates recent developments in AI for LDCT lung cancer screening and is organized around three guiding questions that connect directly to reducing radiologist workload:

• Q1 (native-resolution 2D for transformers/CLIP): When does native-resolution 2D LDCT input help train transformer- or CLIP-based models, and how does the answer depend on the size of the available dataset?
• Q2 (dimensionality for detection/classification): For nodule detection and malignancy classification, is 2D (or 2.5D) input more appropriate than full 3D once computational cost and the limited number of labeled screening cases are accounted for, given that the original NLST randomized 26,722 participants to the LDCT arm, the public TCIA imaging subset contains CT scans from 26,254 subjects, and only 1,060 lung cancers were reported in the LDCT arm?
• Q3 (slice-level attribution): Can identifying which slices contributed most to a prediction provide tangible, workload-reducing value to the radiologist?

We address these questions in the Discussion (Section 8) after synthesizing the evidence.

2. Materials and Methods

We performed a systematic literature search and review following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines [87]. We searched PubMed, IEEE Xplore, Scopus (Elsevier), SPIE Digital Library, and SpringerLink for publications from January 2018 through November 2024. Representative search terms included `lung cancer screening AND (deep learning OR machine learning)”, `lung nodule detection CNN”, `lung CT screening AI”, `transfer learning lung nodule”, and “(NLST OR NELSON) AND lung cancer screening AND (machine learning OR deep learning)”, adapted across databases. We restricted results to English-language studies published from 2018 onward.

The systematic database search was conducted through November 2024, which defines the corpus subjected to formal screening and counting. Because foundation models for radiology advanced rapidly after this cutoff, a small number of high-impact post-cutoff works were additionally tracked through citation chaining and expert knowledge; these are used only to contextualize the discussion and are flagged as post-cutoff additions rather than included in the systematic count. This review was not prospectively registered (e.g. in PROSPERO), which we note as a limitation. To prevent ambiguity, Table 1 states explicitly which reference classes contribute to the systematic count and which are cited for context only.

Records were pooled and de-duplicated; titles and abstracts were screened to remove clearly irrelevant articles. Full texts of the remaining records were assessed against inclusion criteria: a study had to (1) involve low-dose CT lung cancer screening images, and (2) develop, evaluate, or review an AI/ML technique for nodule detection, classification, segmentation, or risk prediction. Both original research and relevant reviews were eligible. Reference lists of key articles were hand-searched for additional studies. In total, 72 studies met the inclusion criteria and form the basis of the thematic synthesis. Beyond these, the reference list also cites landmark screening trials, background and methodological references, screening-guideline statements, and public data resources for context; eleven high-impact works published after the November 2024 search cutoff (foundation models and other 2025–2026 advances surveyed as forward-looking context in Section 7); and one forthcoming companion study (Section 4.6). Consistent with Table 1, these contextual, post-cutoff, and forthcoming references are not part of the systematic count of 72. The flow of records through identification, screening, eligibility, and inclusion is summarized in Figure 1.

Screening and data extraction were performed by the lead author (M.E.H.), with the senior author (F.P.) consulted to adjudicate uncertain inclusion decisions. For each included study we extracted: dataset(s) and cohort type, task (detection, classification, segmentation, or risk prediction), primary input dimensionality (2D/2.5D/3D), model family (CNN, transformer, or hybrid), reported performance metrics and operating points, and validation type (internal vs. external). Explicit eligibility criteria are summarized in Table 2. The database-specific search strategies, the PRISMA 2020 checklist, and a data-extraction sheet for representative included studies are provided as Supplementary Material (Table S1, Supplementary File S1, and Table S3, respectively).

Given the heterogeneity of endpoints and data, no meta-analysis was performed; findings were synthesized thematically (CNN-based detection, malignancy classification, risk prediction, input dimensionality, foundation models, and dataset/metric heterogeneity). To ground the dimensionality discussion in controlled evidence, we additionally summarize an empirical companion study by the present authors (Hoq et al., submitted) [20] that varies only input dimensionality and model family under a matched protocol; consistent with MDPI policy on unpublished data, this study is cited as a forthcoming work, is summarized for context only, and is not counted among the 72 included studies, and its primary data are not re-derived here.

Use of Generative Artificial Intelligence.

During manuscript preparation, the authors used a large language model only for superficial language editing, including grammar, spelling, punctuation, wording, and readability. No generative artificial intelligence tool was used to generate scientific content, formulate the research questions, design the search strategy, select or appraise studies, extract or analyze data, interpret results, create figures or tables, or draw conclusions. All manuscript content was reviewed, verified, and approved by the authors, who take full responsibility for the accuracy and integrity of the work.

3. Results: AI Across the LDCT Screening Workflow

3.1. Overview of Included Studies

Table 3 summarizes representative included studies grouped by task, spanning whole-scan risk prediction, nodule detection, malignancy classification, external validation, and (for post-cutoff context) foundation models. Performance is reported as stated in each primary source; values independently re-verified during preparation are flagged.

3.2. Structured Quality Appraisal

Because performance figures alone can be misleading without attention to study design, Table 4 provides a simplified appraisal of the representative studies along dimensions that matter for screening: external (independent) validation, use of a public dataset, patient-level data splitting, whether sensitivity/specificity (or FROC) is reported, whether calibration is reported, and relevance to the screening reading workflow. The pattern is informative: external validation and calibration reporting are the weakest dimensions across the field, which tempers the high AUCs reported on curated single-institution or diagnostic datasets and motivates the evaluation recommendations in Section 6 and Table 6.

3.3. Automated Nodule Detection

Computer vision algorithms, especially CNN-based models, now achieve excellent sensitivity for lung nodule detection on CT [36]. 2D CNNs on individual slices detect many nodules, but modern systems increasingly leverage 3D information [21]. In a landmark study, Ardila et al. developed a 3D CNN that analyzes whole LDCT volumes; trained on NLST data and evaluated on a held-out set of 6,716 NLST cases, it achieved an AUC of 0.944 for predicting cancer within one year, matching experienced radiologists and detecting some cancers they missed [21]. Single-scan risk models such as Sybil now forecast 1–6-year risk directly from one LDCT without manual annotation, validating across U.S. and Taiwanese cohorts [22].

Many groups report near-radiologist sensitivity while cutting false positives [26]. Liao et al.’s 3D “deep noisy-or” CNN achieved ≈85.6% recall at ∼4 false positives/scan; false-positive reduction approaches using 3D texture/edge features roughly halved baseline false positives [28], and multi-dimensional fusion networks combine 2D and 3D streams for the same purpose [29]. In a routine-clinical-population validation, readers aided by a deep-learning CAD system improved nodule sensitivity from 71.9% to 80.3% with only a small change in the false-positive rate (0.11 to 0.16 per scan) [34], supporting a second-reader role.

Likewise, a separate reader study found that AI-assisted reading improved detection sensitivity, particularly for non-solid nodules [30]. Two-stage pipelines (candidate proposal then false-positive discrimination) are common [37], and pyramid-input designs such as PiaNet reach 93.6% sensitivity at 1 false positive/scan for ground-glass nodules [27]. In practice, approved CADe software highlights suspicious regions for review; user-interface design is critical, since excess false marks erode trust, while precise detectors can reduce reading time by directing attention to a few regions. Comparable gains have been reported across related settings: deep detectors reduce false positives relative to classical machine-learning CAD [38], generalize to multi-lesion detection on chest radiographs [39,40], and have been implemented as modified 3D U-Nets for automated nodule detection [41] and even in preclinical micro-CT [42]. Transfer learning from large natural-image networks remains a common backbone [43], and deep models have been used to confidently reclassify benign nodules so as to curtail unnecessary follow-up [44].

3.4. Nodule Characterization and Malignancy Classification

AI can analyze nodule characteristics to distinguish likely cancers from benign nodules [45]. Causey et al.’s “NoduleX” CNN, trained on >1,000 LIDC-IDRI nodules, reported a nodule-level AUC of ∼0.99 for malignancy and ∼0.95 for nodule-vs-non-nodule discrimination [8]; the attention- and curriculum-based ProCAN reported 98.0% AUC and 95.3% accuracy on LIDC-IDRI [31]. Such near-ceiling figures, however, derive from highly curated datasets with biopsy- or surgery-confirmed ground truth, and real-world performance on small or indeterminate nodules is likely lower. Hybrid models combining CNN features with radiologist-defined features (margin sharpness, lobulation) have reported AUC ≈0.98 while remaining partly interpretable [32]. By assigning objective per-nodule risk scores, such tools could help separate the many benign nodules from the few that warrant work-up, but prospective validation of management impact is still needed. Related deep-learning efforts predict histologic subtype or metastatic potential from CT or PET [46,47], characterize PD-L1 status or immune phenotype [48,49], model longitudinal change with Siamese or recurrent architectures [50,51], improve diagnostic certainty for indeterminate nodules [52], and apply 3D capsule networks to malignancy prediction [53]. Radiomics-based machine learning further links CT phenotypes to prognosis and treatment response [54,55], and deep-learning image reconstruction can preserve nodule conspicuity at reduced dose [56]. More recent (2025) classification work, surveyed in Section 7, includes multitask Swin-transformer characterization of nodules [83] and sequential multi-instance learning across serial screening rounds [84].

3.5. Risk Prediction and Clinical Integration

Beyond individual nodules, AI can estimate overall lung cancer risk [23]. Gao et al.’s co-learning model fused LDCT images with clinical data elements, reaching AUC 0.88 internally and 0.91 on the external Vanderbilt dataset, outperforming imaging-only (0.86) and clinical-only (0.69) models [23]. Temporal models using serial scans stratify risk over time [57], and chest-radiograph models such as CXR-LC (AUC 0.75 vs 0.63 for CMS criteria) could identify high-risk individuals for CT referral [25]. In a screening program, AI risk scores could prioritize worklists and expedite high-risk patients; early real-world deployment, such as the UK study by Murchison et al., detected cancers missed in routine practice without excessive false positives [34], suggesting a “safety-net” role—though prospective validation and reader training remain essential. Since the search cutoff, risk modeling has advanced further (Section 7): 2025 models predict long-term risk from global lung features even in baseline-negative scans [80], with growing emphasis on independent external testing of single-scan risk models [81].

Figure 2 illustrates a generalized conceptual framework that operationalizes the three themes of this review. In this architecture, a quality-control module filters input imaging before it is processed by a self-supervised foundation model backbone. While this backbone can be implemented using either native-resolution 2D slices or full 3D volumetric streams—reflecting the active dimensionality debate in the literature—the core engine utilizes a patient embedding space to retrieve similar historical cancer or benign cases. It then returns a three-way decision (high risk / low risk / indeterminate) accompanied by an auto-generated, case-based explanation. This approach routes AI-flagged low-risk studies toward expedited reassurance and high-risk studies toward prioritized clinical action, aiming to simultaneously reduce false positives and false negatives. Within this broader design space, the authors’ own pipeline serves as one specific instantiation utilizing a 2D-based approach to capture high-resolution slice features.

4. Results: Input Dimensionality and Model Family

Three-dimensional analysis is natural for inherently 3D CT data, and many top-performing systems use 3D CNNs [21]. However, 3D introduces practical misalignments with radiology workflow and statistical challenges that bear directly on Q2.

4.1. Radiologist Reading Habits Versus 3D AI

Radiologists read slice-by-slice and integrate across slices mentally, whereas a 3D CNN may output a whole-volume score not tied to any slice. A bare “60% cancer risk” without localization is hard to act on or trust; useful AI should highlight the nodule, region, or slice driving a prediction. Per-nodule outputs and 3D Grad-CAM-style attributions help, but a fully volumetric workflow remains a paradigm shift not yet routine in practice.

4.2. Interpolation and Preprocessing

3D models often require isotropic voxels and fixed dimensions, so LDCT scans with varying slice thickness are resampled (e.g. $1\times 1\times 1$ mm), introducing smoothing/artifacts that can affect small nodules. To manage memory, developers may downsample or crop, risking missed lesions. This contrasts with the emerging argument that downsampling volumetric CT to low-fidelity 2D slices discards informative grayscale contrast, motivating attention to native-resolution input regardless of dimensionality [12,13].

4.3. Training Cost and Data Requirements

3D networks have far more parameters and need more data. This is problematic given scarce confirmed cancers: the CT arm of the landmark National Lung Screening Trial (NLST), for instance, yielded only 1,060 confirmed lung cancers among its 26,722 participants [2]. A 2D approach treats each slice or patch as a sample (tens of thousands of images), whereas 3D uses one sample per scan/nodule, drastically reducing effective sample size and risking overfitting. A practical compromise is the hybrid 2.5D / 2D–3D approach: wide-net 2D detection followed by limited 3D confirmation [29], balancing sensitivity, stability, and training efficiency.

4.4. Computational Burden

3D models may take 10–30 s per scan (longer on CPU), versus seconds for 2D, which can limit scalability in busy or resource-limited settings. Hardware/optimization advances help—e.g. Lung-CRNet performs 4D-CT registration in under a second [59], and deep networks now enable markerless tumor tracking on simulated 4D-CT [60]—but integration questions (dedicated workstation vs. cloud) remain.

4.5. Segmentation: A Task Where Full 3D Retains Its Advantage

Although classification in the data-limited regime favors 2D/2.5D, segmentation is the task where volumetric context is most valuable and where 3D networks remain the norm, providing an important boundary condition for our Q2 position. Auto-segmentation has advanced rapidly [61], with edge-aware and multidimensional 3D CNNs [62,63], multi-scale deeply supervised 3D U-Nets for lung-tumor delineation [64], and two-stage multitask U-Nets that couple nodule segmentation with malignancy prediction [65]. Transformer-augmented U-Nets increasingly compete with or surpass pure CNNs by modeling long-range dependencies [66,67,68,69], and self-supervised pretraining can reduce the annotation burden that volumetric models otherwise demand [70]. This contrast reinforces our central point: dimensionality should be matched to the task, with full 3D reserved for spatially intensive problems such as segmentation and volumetry, and 2D/2.5D preferred for class-imbalanced screening classification.

4.6. Illustrative Case Study: A Controlled Resource–Performance Comparison (Companion Preprint)

Because this companion study is a preprint that has not yet completed peer review, its results are presented only as an illustrative example and are not used as primary evidence in the systematic synthesis. To isolate input dimensionality from architectural confounds, we summarize a companion study by the present authors (Hoq et al., 2026) [20], available as a preprint and included here for context only (Section 2). That study fixes the training protocol (20 epochs, weighted binary cross-entropy, matched optimization) and varies only input dimensionality—2D (central slice), 2.5D (three orthogonal slices), and 3D (sub-volume)—across a CNN and a Vision Transformer (ViT) on a leakage-free NLST cohort ($n=1,977$; patient-level splits of 1,426/254/297) prepared with a lung-focused quality-control pipeline [58]; LIDC-IDRI provided supporting weak labels.

As summarized in Table 5, the 2.5D CNN achieved the best discrimination (ROC-AUC 0.682, 95% CI [0.546, 0.799]) with stable operating behavior. The 2D CNN was overly conservative; the 3D CNN showed threshold instability (sensitivity 0.05–0.50 between default and validation thresholds). Transformers required ≈3× the GPU memory and frequently collapsed: the 3D ViT produced zero sensitivity (all-negative) despite ROC-AUC 0.589, and the 2.5D ViT showed extreme threshold sensitivity. The pattern reproduced on LIDC-IDRI, where several higher-capacity models collapsed to all-positive behavior (specificity 0) despite moderate AUC. Two lessons recur in this review: ROC-AUC alone can mask clinical unusability (acceptable ranking yet extreme sensitivity–specificity imbalance), and input dimensionality governs not only performance but failure mode, with full 3D and transformer configurations most prone to degeneracy in the data-limited, class-imbalanced screening regime.

4.7. Native-Resolution 2D Input and Transformer Data Efficiency

ViTs lack the inductive biases of CNNs (locality, translation equivariance) and need large-scale data [71]. In screening applications, this data dependency frequently manifests as the training instability and representation collapse illustrated in Section 4.6. Two lines of evidence indicate when native-resolution 2D input helps a transformer-/CLIP-based model. First, self-supervised 2D foundation models shift the representation burden off scarce labels: a feasibility study with native-resolution 2D RAD-DINO embeddings achieved strong classification because the heavy lifting occurs during large-scale pretraining [12]. Second, recent radiology foundation-model work argues that preserving native resolution (rather than downsampling to low-fidelity slices), with large-scale pretraining, yields strong, data-efficient performance and improves long-horizon risk prediction [13]. Task-agnostic 3D pretraining is an alternative route to data efficiency when ample unlabeled volumes exist [72], but does not negate the central point for small labeled data: native-resolution 2D plus pretraining is the more reliable recipe for transformer/CLIP models, with the benefit shrinking as the labeled dataset grows.

4.8. Slice-Level Attribution as a Workload-Reducing Aid

Because reading is slice-by-slice, the most actionable output beyond a per-case score is an indication of which slices and regions drove the prediction. Per-slice/per-region attribution (attention, Grad-CAM-style saliency, slice-importance scoring) turns an opaque output into a triage aid, e.g., directing the reader to the few slices that matter reduces images requiring close scrutiny and shortens reading time. The 2.5D paradigm aligns naturally with this, since its contributions are attributable to identifiable slices, whereas volumetric models require post-hoc reconstruction. Current interpretability work emphasizes nodule-level heatmaps over reader-facing slice triage [73], which remains a concrete gap. Reliable slice-level attribution also requires preserving per-case predictions [20].

5. Results: Foundation Models

Foundation models—large, broadly pretrained models adaptable to many tasks—are maturing in medical imaging but face several limitations relevant to screening [11].

5.1. Domain Mismatch and Representation

General vision models trained on natural images transfer imperfectly to grayscale CT; domain-specific pretraining helps. For example, MIS-FM, pretrained on 110,000 unlabeled 3D CT scans, outperformed prior segmentation methods after fine-tuning [74], and CT-specific self-supervised approaches—native-resolution 2D [12] and task-agnostic 3D pretraining [72]—target this gap. Cross-domain experience with transfer learning in other tumor types underscores both the promise and the fragility of borrowed representations [15,16].

5.2. Data and Computational Requirements

Hierarchical transformers (e.g. Swin) have tens of millions of parameters [75], and the scarcity of labeled screening data pushes the field toward self-supervision; contrastive transformers for nodule detection, for instance, improve sparse-label performance [35]. Because no single center holds enough labeled cancers, distributed and federated training across institutions is increasingly proposed to assemble the needed scale without pooling patient data [76], and cloud-based deployment can offload inference cost [17].

5.3. Interpretability and Customization

Large models are less interpretable than task-specific networks, and fine-tuning risks catastrophic forgetting; explainability for screening therefore remains an active need [73].

5.4. Current Performance and Examples

No foundation model has clearly displaced specialized CNNs across all lung tasks, though transformer skip-connection designs improve segmentation [66]. Within the formal search window, cross-domain deep-learning pipelines (e.g., for chest-radiograph triage during the COVID-19 pandemic) demonstrate how rapidly such systems can be assembled and validated when data are available [77]. While these baseline architectures show strong promise, the trajectory has shifted toward specialized radiology and multimodal foundation models optimized specifically for screening workflows; these post-cutoff advancements and their data-efficient gains are analyzed in detail in Section 7.

5.5. Multi-Modality and Few-Shot Adaptation

Linking imaging with text reports enables human-readable explanations and few-shot adaptation [78], an encouraging but still-maturing direction that is particularly relevant to the reader-facing, explanation-oriented integration envisioned in Figure 2.

6. Results: Dataset and Evaluation Heterogeneity

A major obstacle to assessing progress is heterogeneity in datasets and metrics [33]. Public datasets differ substantially in clinical utility: for instance, while the widely used LIDC-IDRI dataset is fully accessible via The Cancer Imaging Archive (TCIA) [79], it represents an older diagnostic cohort whose acquisition parameters do not reflect the spatial resolution and capabilities of modern low-dose CT scanners. Meanwhile, NLST has a limited public release, and the Kaggle DSB 2017 dataset lacked granular annotations. Few studies test on independent external data; Jacobs et al. found top algorithms reached AUC 0.877 to 0.902 versus radiologists’ 0.917, which is close but does not significantly exceed human readers [33]. Ground-truth definitions vary, such as detailed multi-reader annotation versus confirmed-cancer labels, as do operating points, making comparison difficult without common FROC analysis. Metric definitions, specifically patient- versus nodule-level sensitivity, are not interchangeable, and at a realistic 1 to 2% prevalence even 95% sensitivity with 80% specificity yields more false positives than true positives. As the companion analysis underscores, reporting ROC-AUC without operating-point behavior can hide degenerate collapse [20]. Ultimately, a modern benchmark explicitly linking detection to clinical outcomes in a contemporary screening cohort is still needed; models tuned on older cohorts also generalize imperfectly to modern European protocols like NELSON-type populations. Encouragingly, recent (2025) studies increasingly report independent external validation, including of commercially available detectors in new national screening cohorts [85], and updated European recommendations now explicitly call for deep-learning-assisted nodule detection and growth assessment [86].

7. Recent Developments (2025–2026)

The systematic search closed in November 2024. Because the field has moved quickly since, we summarize here, strictly as forward-looking context outside the systematic count (Table 1), several 2025–2026 studies that reinforce the conclusions of this review.

Risk prediction has extended beyond visible nodules. ScreenLungNet combined multiple-nodule and global lung-parenchyma features to predict three-year lung-cancer risk from a single LDCT (AUC 0.93–0.94) and retained useful accuracy even in baseline-negative participants (AUC 0.87), addressing the interval and post-screening cancers that lack a visible baseline nodule [80]. In parallel, independent external testing of single-scan deep-learning risk models on LDCT has begun to close the external-validation gap emphasized in Section 6 [81].

Foundation and transformer models continued to mature in directions consistent with Q1. A multimodal, multitask foundation model trained for lung cancer screening reported coverage across detection, classification, and risk tasks from a single pretrained backbone [82], while a multitask Swin transformer jointly classified and characterized pulmonary nodules [83]. These results reinforce our position that large-scale pretraining, rather than raw architectural capacity, is what makes transformers viable in the data-limited screening regime. Sequential multi-instance learning across serial screening rounds further exploited longitudinal structure for classification [84].

Finally, the evidence base has shifted toward external validation and clinical translation. External validation of commercially available deep-learning nodule detection in a Japanese LDCT screening population [85], together with updated European practice recommendations that now explicitly call for deep-learning–assisted nodule detection and volumetric growth assessment [86], signals movement from algorithm development toward deployment—while reaffirming, as this review argues, that prospective validation and standardized, operating-point–aware evaluation remain the principal bottlenecks.

8. Discussion

We answer the three guiding questions, each framed by the goal of reducing radiologist workload, then address real-world integration and the limitations of this review.

8.1. Q1: Native-Resolution 2D for Transformer/CLIP Models Is Helpful, Conditional on Dataset Size

The evidence supports a qualified “yes.” Because transformers lack CNN inductive biases and are data-hungry [71], training them—especially in 3D—on the small labeled cohorts typical of screening tends to be unstable (Table 5) [20]. Native-resolution 2D input becomes advantageous when paired with large-scale self-supervised or pretrained representations that move the data burden off scarce labels [12,13]. The benefit is largest when labeled data are modest and shrinks as the dataset grows; for the small-to-moderate datasets characterizing most screening research, native-resolution 2D plus pretraining is the more dependable recipe.

8.2. Q2: For Detection and Classification, 2D/2.5D Is the Better Trade-Off than Full 3D in the Screening Regime

The data-availability argument is decisive: with even NLST offering only ∼1,000 cancers among ∼26,000 CT-arm participants [2], 3D models that consume one sample per scan are starved of examples, whereas 2D/2.5D multiply effective training data by treating slices as samples (Section 4.3). This is corroborated by published work in which hybrid 2D–3D pipelines capture most of the benefit of volumetric context without the full 3D penalty [29]. Consistent with this published evidence—and offered only as an illustration, not as primary evidence—our companion comparison found the 2.5D CNN best with stable behavior, the 3D CNN threshold-unstable, and transformer/3D configurations prone to collapse, all at higher GPU cost for volumetric/ViT variants [20]. The caveat is task-dependence: full 3D remains preferable for spatially intensive tasks such as segmentation and volumetry. For classification in class-imbalanced screening, 2D and especially 2.5D are the more reliable defaults.

8.3. Q3: Slice-Level Attribution Offers Substantial, Underexploited Value

Because reading is slice-by-slice, slice-level attribution is the most direct route by which AI reduces workload—turning a per-case score into a triage signal that points the reader to the few slices that matter (Section 4.8). The 2.5D paradigm is naturally suited to this; volumetric models require post-hoc reconstruction. Current work emphasizes nodule-level heatmaps over reader-facing slice triage [73], leaving a tractable, high-value gap: standardized, validated slice-level attribution presented inside the PACS reading interface would convert model accuracy into measurable workflow efficiency.

8.4. Real-World Integration Challenges

Translation requires regulatory clearance (which is often more straightforward for narrow CADe tasks than holistic risk prediction), seamless PACS/DICOM integration with real-time performance, and human-factors design so that AI marks assist rather than distract the clinical reader. These practical hurdles are closely mirrored by broader systemic concerns; medico-legal responsibility, reimbursement pathways, out-of-distribution detection, equity monitoring across demographic subgroups, and patient transparency all remain substantial barriers to routine deployment. To address these multi-faceted requirements systematically, emerging frameworks like the FUTURE-AI international consensus guidelines provide a structured blueprint for achieving trustworthy and clinically deployable healthcare AI [88]. Ultimately, a phased deployment, where early adopters collect prospective clinical impact data alongside professional-society guidance as evidence accrues, represents the most viable path forward.

8.5. Clinical Implications for LDCT Lung Cancer Screening

Translating these findings into screening practice suggests several concrete implications. First, AI for screening should not be optimized for ROC-AUC alone; in a population with ∼1–2% cancer prevalence, the operationally decisive endpoints are reductions in unnecessary recalls (false positives) and in missed cancers (false negatives), so calibration and operating-point behavior must be reported and optimized. Second, 2D/2.5D models—being lighter and more stable in the data-limited regime—may be more readily deployable in screening clinics with limited compute, widening equitable access. Third, slice-level attribution aligned with how radiologists read could reduce reading burden by guiding the reader to the few relevant slices, a tractable efficiency gain. Fourth, AI outputs should be integrated with established decision structures—Lung-RADS categorization, PACS/DICOM workflow, and multidisciplinary management pathways—rather than presented in isolation. Finally, prospective, multi-center validation that measures management impact (time-to-read, recall rates, stage shift) is needed before clinical adoption; current evidence, while promising, is largely retrospective.

8.6. Limitations of this Review

The systematic search closed in November 2024; high-impact 2025–2026 developments are surveyed separately as flagged, post-cutoff context (Section 7) and are not part of the systematic count, but a fully updated systematic search may surface additional studies. The review was not prospectively registered, and English-only inclusion may introduce selection bias. Because screening and data extraction were performed by one primary reviewer with senior-author adjudication, selection bias in study inclusion cannot be fully excluded. Heterogeneity of endpoints precluded meta-analysis, so synthesis is narrative. Finally, the companion empirical study is the authors’ own and is not yet peer-reviewed; it is cited as forthcoming and used only to illustrate, not to establish, the dimensionality argument.

9. Conclusions and Future Directions

AI can augment radiologist performance across LDCT screening: CNN detectors reach high sensitivity for small nodules, attention-based and fusion classifiers assess malignancy on par with experts, and imaging-plus-clinical risk models refine patient selection, all enabled by public resources such as LIDC-IDRI and NLST [79]. Our synthesis reaches three positions. First, native-resolution 2D input is a dependable basis for transformer-/CLIP-based models when labeled data are limited and large-scale pretraining is available, conditional on dataset size. Second, for nodule detection and malignancy classification in the data-limited, class-imbalanced screening regime, 2D and especially 2.5D inputs may offer a more favorable performance–stability–cost trade-off than full 3D, which retains its advantage for segmentation and volumetry. Third, slice-level attribution—aligned with the 2.5D paradigm and with how radiologists read—is an underexploited, tractable route to reducing reading burden. Future work should prioritize prospective multi-center validation, standardized evaluation that reports operating-point behavior alongside ROC-AUC, and reader-facing slice-level interpretability integrated into clinical workflow, so that algorithmic accuracy translates into earlier detection, fewer unnecessary procedures, and a lighter load on radiologists. Table 6 summarizes the principal gaps and a recommended research agenda.