Role of Imaging Techniques in Ovarian Cancer Diagnosis: Current Approaches and Future Directions

1. Introduction

In 2020, ovarian cancer (OC) ranked eighth on the list of the most diagnosed malignant diseases in the world involving approximately 314000 women, moreover being eighth in the overall ranking of cancers most associated with mortality with more than 207000 deaths [1]. According to some studies there is a trend related to an increase in the incidence of gynecological cancers [2]. The most common malignant subtype of OC is high-grade serous carcinoma (≈70%). The prognosis of ovarian cancer tends to be unfavorable, and it is related to the inherent malignancy of the neoplasm itself, the absence of specific signs and symptoms that allow early diagnosis, and the lack of valid mass screening. Since ovarian cancer has the tendency to present clinically at an advanced stage (stage III and IV according to FIGO staging 2021 in 70-80% of cases), overall 5-year survival is estimated to be around 46%, effectively making ovarian cancer one of the most lethal cancers [1]. Staging is surgical and based on the FIGO system. Standard treatment includes cytoreductive surgery and platinum-taxane chemotherapy. Neoadjuvant chemotherapy is used when primary surgery is not feasible. Targeted therapies such as PARP inhibitors (for BRCA-mutated tumors) and bevacizumab (anti-angiogenic) have improved outcomes. HIPEC and immunotherapy are under evaluation. Radiotherapy has a limited role, mainly in palliative settings [3,4,5].

2. Imaging in Ovarian Carcinoma

2.1. Ultrasonographic Imaging

Ultrasound is the first-line imaging modality in the evaluation and characterization of adnexal masses in both healthy and high-risk patients. The sensitivity and specificity of ultrasonography for the characterization of adnexal masses are influenced by several factors, including the operator’s experience, the facility where the ultrasound is performed, the quality of the equipment, and the type of patient and the mass itself. Clinical studies and researches have suggested average values of sensitivity and specificity that may vary, with sensitivity typically ranging from 70% to 90% and specificity from 70% to 95% [6,7].

Ultrasonography relies on the morphologic features of ovarian lesions, which, when assessed in conjunction with menopausal status and CA-125 levels, allow an estimate of the risk of malignancy. The proficiency of the operator, along with the reputation and resources of the investigative centre, plays a critical role in determining the accuracy of diagnosis and the effectiveness of prognosis for the patient. For this reason, the goal over the years has been to develop models for stratifying the risk of malignancy of ovarian lesions that would be simple to apply, reproducible, and effective. Over the years various attempts have been made to address this issue, including the Risk of Malignancy Index (RMI), Risk of Ovarian Malignancy Algorithm (ROMA) and Logistic Regressions 1 and 2 (LR1 and LR2). The most recent risk stratification models are the ADNEX model and the ORADS-US model.

2.1.1. IOTA ADNEX Model

The IOTA-ADNEX (International Ovarian Tumor Analysis-Assessment of Different Neoplasias in the Adnexa) model was published in 2014 with the aim of classifying ovarian lesions on ultrasound, and it proved to be more effective than previously used models. The ADNEX model aims not only to predict the type of lesion (benign or malignant) but also to describe the “stage” of disease (borderline, stage I, stage II-IV, metastatic), leading to the early detection of suspicious lesions and, consequently, the reduction of mortality and morbidity [8]. This model takes into account nine factors: age (in years), serum CA-125 levels (U/ml), type of center (reference/non-reference in oncology), maximum diameter of the lesion (mm), number of papillary projections (0, 1, 2, 3 or >3), maximum diameter of the major solid portion (mm), more or less than 10 cystic loculi (yes/no) and ascitic effusion (yes/no). The various results are then translated both graphically and numerically into risk of lesion malignancy (%), in which the ADNEX model was found to have high specificity.

2.1.2. O-RADS US

The Ovarian-Adnexal Reporting and Data System (O-RADS) US is a risk stratification and management system for ovarian and adnexal lesions based on ultrasound characteristics of lesions. Published in 2018, its aim is to create a common language to minimize ambiguity and improving the interpretation of ovarian masses [6,9].

The O-RADS US system is based on a standardized lexicon, which allows ovarian lesions to be placed into six different categories (0-5), each of which is associated with different risk of malignancy of the lesion:

• O-RADS 0: incomplete evaluation
• O-RADS 1: physiological findings
• O-RADS 2: lesions almost certainly benign (<1% risk of malignancy)
• O-RADS 3: lesions with low risk of malignancy (1-10%)
• O-RADS 4: lesions with intermediate risk of malignancy (10-50%)
• O-RADS 5: lesions with high risk of malignancy (>50%)

Categorization is done by evaluating features such as size, margins, number of loculi, solid components, color-score (1-4 at echo-colordoppler evaluation), number of papillary projections, ascites and peritoneal implants (Table 1, Figure 1).

2.2. CT Imaging

The role of CT imaging is limited in the characterization of adnexal masses; however, it is critical in loco-regional and especially in distant staging of disease. This imaging method boasts of wide availability, relatively low cost and short acquisition time. The CT acquisition protocol involves obtaining scans of 1–2 mm thickness of the neck, thorax, abdomen, and pelvis after the injection of iodinated contrast medium. Venous phase images should be acquired 70-80 seconds after contrast injection, followed by MPR in both the coronal and sagittal planes (Figure 2). There is still no consensus on the use of oral contrast medium administered with the goal of better evaluation of bowel loops and peritoneal carcinosis implants. Oral contrast medium can be positive (iodinated mdc) or negative (water).

CT examination allows us to describe the relationships of the mass to surrounding structures (although with less specificity than with MRI), to identify peritoneal localizations of disease and any lymph node and parenchymal metastases in the abdominal and extra-abdominal locations. Contrast-enhanced CT (CE-CT) has a reported staging accuracy for ovarian cancer of up to 94%, allowing a better prediction of successful surgical cytoreduction. It is important to emphasize the importance of identifying and describing all peritoneal localizations, firstly to assess the operability (or not) of the patient, and then for proper surgical planning. The most recent ESUR guidelines 2025 [10] highlight the importance of a structured report to describe disease locations not only at diagnosis but also in follow-up. In this way, the radiologist not only identifies and describes areas where disease is present, but also includes in the report the potential resectability or unresectability of disease sites that may limit successful R=0 after surgery, particularly in cases of mesentery root involvement, diffuse carcinomatosis on the surface of the small intestine (resection of which would result in short bowel syndrome), diffuse involvement/infiltration of the stomach, duodenum, head of the pancreas, and finally involvement of the celiac axis and its branches (celiac lymph nodes may be removed) [11,12,13].

2.3. MRI Imagingg

Although several US diagnostic scores that effectively stratify the risk of malignancy in ovarian lesions with high sensitivity and specificity have been reported in the literature (see Chapter 2.1), up to 30% of adnexal masses remain indeterminate on US, thereby posing significant challenges for treatment planning [14,15]. In this context, Magnetic Resonance Imaging (MRI) with contrast medium represents the imaging method of choice for the characterization of “indeterminate” ovarian lesions. In fact, MRI provides more accurate diagnoses of lesions that cannot be characterized on US, thereby reducing the level of suspicion and consequently the number of unnecessary surgeries on benign lesions in asymptomatic patients, achieving a PPV of 71% and a NPV of 98% and diagnostic accuracy values ranging from 83% to 93% for the diagnosis of malignancy [14,16,17,18,19]. In addition, the inherent high contrast resolution and the possibility to characterize tissues allows the better characterization of the lesion; this is compounded by the possibility, due to the method’s multiparametricity and multiplanarity, to investigate all portions of the lesion, some of which may not be ultrasound-investigable trans-vaginally and/or trans-abdominally [20]. MRI also plays a key role in the characterization of voluminous lesions of doubtful adnexal origin at US evaluation alone, which up to 10% of cases ultimately being of non-annexal relevance, and therefore reclassified [14].

As in ultrasound diagnosis, a common language has been sought in MRI characterization of adnexal masses that is reproducible, easily understandable, and comprehensive for describing adnexal masses and estimating their risk of malignancy. The first in this regard was the ADNEX-MR model proposed by Thomassin-Naggara et al in 2013 [21].

2.3.1. ADNEX-MR Scoring System

The aim of the “ADNEX-MR” Scoring System model is to build a reproducible preliminary validation method dedicated to the use of pelvic MR in the characterization of adnexal masses which resulted “indeterminate” at ultrasound evaluation. In this study, the Auhors made use of 1.5T MRI scanner with “phased-array” coils, after treatment of the patients with intravenous spasmolytic drug for better image resolution (in particular, glucagon), using the following protocol: axial and sagittal non-“fat-sat” T2 - TSE sequences, T1 gradient-echo sequences with and without fat saturation, DWI in the axial plane with b levels up to 1000 sec/mm². Sequential T1 GRE perfusion sequences were acquired during IV administration of the contrast medium with a temporal resolution interval of 2.4 seconds for a total of 320 seconds, starting 10 seconds before contrast medium pier injection (2 ml/sec); finally a late axial or sagittal T1 gradient sequence. The purpose of the perfusion study was the comparison of myometrial enhancement with that of the solid tissue in the adnexal lesion. In this context, solid tissue is defined as the presence of vegetations, solid areas, or irregular thickening of the septa or walls [22]. Specifically, two ROIs (region of interest) were placed, one in the outer myometrium and the other in the portion of solid tissue with more enhancement, with the purpose of making a comparison between two time-signal curves. A gradual increase in the signal (enhancement) of solid tissue in the absence of a well-defined peak corresponds to a type 1 curve. A moderate initial increase in the solid tissue signal relative to that of the myometrium followed by a plateau is considered a type 2 curve. Finally, a type 3 curve is defined as such when the peak of solid tissue enhancement in the lesion is earlier and steeper than that of the myometrium.

Despite the many features analyzed during the evaluation of ovarian lesions, the criteria taken into consideration to classify ovarian lesions are as follows: presence or absence of solid tissue, solid tissue signal in T2w sequences and in diffusion (DWI) sequences, type of enhancement curve, and presence of peritoneal implants (Table 2).

According to the above-described ADNEX model, ovarian lesions can be classified with the following scores:

• Absence of ovarian lesions
• Benign lesions: lesions with homogeneous content (serous, blood, fat) and the absence of wall enhancement and/or with hypointense solid tissue signal in T2 sequences and at high b values
• Probably benign lesions: absence of solid tissue or solid tissue with type 1 enhancement curve
• Indeterminate lesions: presence of solid tissue with type 2 curve
• Likely malignant lesions: type 3 enhancement curve and presence of peritoneal implants

The ADNEX model was found to predict the risk of malignancy with more than 90% accuracy. According to the results obtained, the type 3 curve is the most predictive criterion for malignancy, along with the presence of irregular septa and peritoneal implants. Conversely, in Score 2, the risk of malignancy is extremely low (<2%) [21].

2.3.2. O-RADS MRI Score

The O-RADS (Ovarian-Adnexal Reporting and Data System) MRI score represents the most recent and widely accepted system for stratifying the risk of malignancy in ovarian lesions, developed by an international group of experts from various disciplines. By estimating the probability of malignancy in ovarian lesions, it aims to increase accuracy in the characterization of adnexal lesions, enhance interdisciplinary communication, and consequently promote optimal management in these patients [16]. The O-RADS MRI system represents an evolution of the ADNEX MR model [21] described in the previous paragraph, and it benefits from an overall diagnostic accuracy of 92%, associated with a sensitivity of 93% and specificity of 91%, PPV of 71% and NPV of 98% [14,15,23]. The first part of the O-RADS model building process involves the use of correct and standardized terminology. In this context, firstly a lesion is defined as a non-physiologic relief; then the features and definitions of iconographic findings are well distinguished, for instance the type of signal (homogeneous/inhomogeneous and hypointense/intermediate/hyperintense), the type of lesion (cystic/solid), the type of fluid (simple/hemorrhagic/proteinaceous/fat), solid components (papillary projections/mural nodules/irregular patches/solid portions) with solid tissue being defined as tissue exhibiting enhancement, non-solid tissue (clots/debris/fibrin/etc.), and so on. For a more comprehensive evaluation of terminology, please refer to the following articles [16,24].

Acquisition protocol

The examination protocol involves the acquisition of non-FS sagittal T2w sequences (≤ 4 mm), axial T2w sequences (≤ 3 mm), in-phase and out-phase T1-weighted sequences (≤ 4 mm), DWI in the axial plane (≤ 4 mm; b value > 1000 sec/mm²), and finally post-contrast T1-weighted sequences. In this regard, along the lines of the ADNEX model, a dynamic study (DCE) T1-weighted sequences characterized by a resolution time of 15 seconds and a slice thickness of 3 mm is proposed, with acquisition beginning 30 seconds after contrast medium injection and lasting for 4 minutes. This study allows calculation of signal/time curves by placing ROIs on the solid tissue of the lesion and on the myometrium, being careful not to place the latter on the vessels. Specifically, the signal/time curve can be of three types when compared with the myometrial curve (Figure 3):

• Type1(low risk): slower increase in signal intensity (enhancement) of solid tissue than that of the myometrium, without a peak or plateau
• Type 2 (intermediate risk): moderate initial increase in solid tissue signal with a slower or equal slope to that of the myometrial, followed by a plateau
• Type 3 (high risk): pronounced signal increase with an early peak compared with that of myometrium

In case a perfusion study (DCE with resolution < 15 sec) cannot be acquired, this can be replaced by acquiring an ultrafast grandient-eco at 30-40 sec after contrast injection (thickness ≤ 3 mm), thus allowing a different type of comparison with myometrium. Additionally, in patients who have undergone a hysterectomy, the radiologist can evaluate the morphology of the signal/intensity curve concerning the solid portions with contrast enhancement, taking into account the curve morphology (with or without a plateau).

CLASSIFICATION

O-RADS MRI Score 0

Lesions that cannot be properly evaluated on MRI examination fall into this category; for example, when some portions of the lesion are not included in the study volume, when all sequences could not be acquired, or when artifacts compromise the quality of the examination.

O-RADS MRI Score 1

Findings of normality. In premenopausal women, follicles, corpora lutea, and hemorrhagic cysts up to a maximum size of 3 cm are classified as O-RADS 1; therefore, these findings are not considered pathological. In postmenopausal women, some follicles may still be present in the ovarian parenchyma, and if the radiologist defines these findings in the “normal range”, they will also receive a score of 1, thus not considering them lesions.

O-RADS MRI Score 2 (risk of malignancy < 0.5%)

In both pre- and postmenopausal women, it is important to apply this score only in those formations considered to be lesions, as follows:

Unilocular cystic lesions with homogeneous fluid content, without wall enhancement and without solid tissue

• Unilocular cysts with uncomplicated fluid or endometrioid content, with mild wall enhancement but without intracystic solid portions (e.g., endometrioma)
• Adipose-content lesions without intralesional solid tissue; in this case, the adipose tissue shows hyperintense signal in T2- and T1-weighted sequences, with signal drop in fat-sat sequences. It is necessary to distinguish Rokitansky nodule, which may show enhancement, from solid tissue (e.g., mature teratoma).
• Solid lesions with homogeneously hypointense signal both in T2w and DWI sequences at high b-values, regardless of the type of enhancement after mdc (e.g., ovarian fibroma)
• Fallopian tubes dilated by simple fluid, with mild and subtle wall enhancement in the absence of solid tissue
• Para-ovarian cysts at any type of fluid content, which may or may not show wall enhancement, without intralesional solid tissue.

O-RADS MRI Score 3 (risk of malignancy ≈ 5%)

• Unilocular cysts (proteinaceous/hemorrhagic/mucinous fluid content) with wall enhancement but without solid tissue
• Multilocular cysts at any type of fluid content, with thin septa that may show enhancement and absence of solid tissue
• Lesions with solid tissue (excluding solid lesions described in score 2) that show low-risk enhancement curve (type1).
• Fallopian tubes with non-simple fluid content, thickened walls, no solid tissue.

O-RADS MRI Score 4 (risk of malignancy ≈ 50%)

• Lesions with solid tissue (excluding solid lesions described in score 2) showing type 2 enhancement curve (intermediate risk; Figure 4)
• Solid lesions showing enhancement < myometrial at 30-40 seconds, if perfusion study is not available
• Lesions with lipid components and solid tissue with enhancement

O-RADS MRI Score 5 (risk of malignancy ≈ 90%)

• Lesions showing a type 3 enhancement curve (high risk)
• Solid lesions showing enhancement > myometrium at 30-40 seconds, if perfusion study is not available
• Lesions associated with the presence of peritoneal implants and/or secondary disease localization (Figure 5)

The primary advantage of the O-RADS MRI Score is the high NPV for malignancy (about 98%) in lesions classified as Score 2 or 3, thus allowing the radiologist to express himself or herself in those terms with a high degree of confidence [14]. Other strengths are the easily applicable and reproducible method, the lexicon, and the extensive validation in the clinical setting [23,25]. These improvements result in a reduction of unnecessary surgeries and improvement of therapeutic planning for suspected ovarian lesions, particularly in young premenopausal women who may benefit from conservative treatment. According to Thomassin-Nagarra et al. the majority of errors in O-RADS MRI grading are attributable to misinterpretation of solid tissue and incorrect determination of the mass’s origin [23].

3. New Perspectives and AI in Ovarian Cancer

3.1. O-RADS MRI/ADC Score

Several studies have attempted to implement the O-RADS MRI Score, particularly category 4, where the PPV range for malignancy is particularly wide (5-90%), highlighting the potential usefulness of the apparent diffusion coefficient (ADC) map of both cystic and solid components [26,27,28]. These studies have shown how the signal intensity in diffusion sequences, along with the corresponding values in ADC maps, can change the radiological classification according to the traditional O-RADS model. This adjustment allows the upgrading or downgrading of lesions between scores 3-4 and 4-5, thereby reducing the large variability inherent in category 4, as described in the “O-RADS MRI/ADC Score” proposed by Manganaro et al. [28]. This investigation also demonstrated a significant correlation between the values in the ADC map and the histologic variant of the tumor, in particular how the mean ADC values of borderline tumors are higher than in low-grade and high-grade serous carcinomas.

3.2. NON- CONTRAST- MRI Score

A retrospective study by Sahin et al [29] evaluated the possibility of characterizing and stratifying the malignancy risk of adnexal lesions without the aid of contrast medium perfusion. This was achieved by constructing a 5-point score based on morphologic and qualitative characteristics (DWI), redefining the term “solid tissue” as “tissue component showing intermediate signal in T2w sequences, low signal in T1w and diffusion restriction in DWI(ADC) sequences” (Table 3; Figure 6). The “Non-contrast MRI Score” model has proven to be a reliable and reproducible tool with high diagnostic accuracy (94.2 %), excellent levels of sensitivity (84.9 %) and specificity (95.9 %), and significant inter-reader agreement. It, therefore, potentially represents a viable alternative to the O-RADS MRI score in cases where contrast medium administration is not recommended, such as in cases of severe allergic reactions to gadolinium-based contrast medium, pregnancy status and severe nephropathy. The main limitation is the lack of external validation by a prospective, multicenter study; however, the latter is in an active phase and shows promising preliminary results.

3.3. PET-MRI

Although PET-MRI has been clinically available from a decade, its adoption remains limited compared to the widespread use of [18F] FDG PET-CT. This is due to a lack of standardised imaging protocols and workflow harmonisation, as well as costs, which hinder broader clinical use [10]. In ovarian cancer, PET-MRI has demonstrated a better diagnostic performance over PET-CT and diffusion-weighted MRI (DWI-MRI) in peritoneal staging at initial diagnosis (p = 0.001), though no significant difference was found after chemotherapy [30,31]. Preliminary evidence also supports PET-MRI’s utility in detecting recurrence of gynecologic malignancies; however improving diagnostic randomized studies are needed to fully define its added value in routine clinical practice [32].

Table 4. Comparison of different imaging modalities in ovarian cancer.

Modality	Main Uses	Sensitivity / Specificity	Strengths	Limitations
Transvaginal Ultrasound (TVUS)	First-line evaluation of adnexal masses. Assessment of morphology and vascularization.	Sensitivity: ~85% / Specificity: ~90% (for mass characterization)	Widely available, non-invasive, no radiation. High-resolution for pelvic organs.	Operator-dependent. Limited for staging and evaluation of peritoneal spread.
Computed Tomography (CT)	Staging (especially peritoneal, nodal, and distant metastases). Preoperative planning.	Sensitivity: ~70–85% / Specificity: ~90% (for staging)	Good overview of abdomen and pelvis. Useful for surgical planning and monitoring recurrence.	Limited soft tissue contrast. Poor detection of small peritoneal implants (<1 cm).
Magnetic Resonance Imaging (MRI)	Characterization of indeterminate masses. Local staging. Evaluation of complex cystic lesions.	Sensitivity: ~93% / Specificity: ~88–98%	Excellent soft tissue contrast. No radiation. Functional imaging (DWI) adds value in lesion analysis.	More expensive, time-consuming. Contraindicated in patients with certain implants.
FDG-PET/CT	Detection of recurrence, metastases. Useful in equivocal cases.	Sensitivity: ~80–95% / Specificity: ~90–95% (in recurrence)	Functional and anatomical data. High sensitivity in detecting active disease and distant spread.	Limited role in primary diagnosis. False positives in inflammation/endometriosis. Costly.
PET/MRI (emerging)	Research setting. Combines metabolic and high-resolution anatomic data.	Under investigation	Potentially best of both PET and MRI. Promising for advanced imaging and radiomic studies.	Limited availability. High cost. Not yet widely implemented in clinical practice.

Table 4. Comparison of different imaging modalities in ovarian cancer.

Modality	Main Uses	Sensitivity / Specificity	Strengths	Limitations
Transvaginal Ultrasound (TVUS)	First-line evaluation of adnexal masses. Assessment of morphology and vascularization.	Sensitivity: ~85% / Specificity: ~90% (for mass characterization)	Widely available, non-invasive, no radiation. High-resolution for pelvic organs.	Operator-dependent. Limited for staging and evaluation of peritoneal spread.
Computed Tomography (CT)	Staging (especially peritoneal, nodal, and distant metastases). Preoperative planning.	Sensitivity: ~70–85% / Specificity: ~90% (for staging)	Good overview of abdomen and pelvis. Useful for surgical planning and monitoring recurrence.	Limited soft tissue contrast. Poor detection of small peritoneal implants (<1 cm).
Magnetic Resonance Imaging (MRI)	Characterization of indeterminate masses. Local staging. Evaluation of complex cystic lesions.	Sensitivity: ~93% / Specificity: ~88–98%	Excellent soft tissue contrast. No radiation. Functional imaging (DWI) adds value in lesion analysis.	More expensive, time-consuming. Contraindicated in patients with certain implants.
FDG-PET/CT	Detection of recurrence, metastases. Useful in equivocal cases.	Sensitivity: ~80–95% / Specificity: ~90–95% (in recurrence)	Functional and anatomical data. High sensitivity in detecting active disease and distant spread.	Limited role in primary diagnosis. False positives in inflammation/endometriosis. Costly.
PET/MRI (emerging)	Research setting. Combines metabolic and high-resolution anatomic data.	Under investigation	Potentially best of both PET and MRI. Promising for advanced imaging and radiomic studies.	Limited availability. High cost. Not yet widely implemented in clinical practice.

3.4. Artificial Intelligence (AI)

Artificial intelligence (AI) is taking an increasingly important role in medical diagnostics, including the imaging of ovarian cancer. This technology is used to improve the accuracy, speed, and efficiency of diagnostic processes, particularly in the analysis of medical images obtained by computed tomography (CT), magnetic resonance imaging (MRI), or ultrasound (US) [33]. It is now well known that high-grade serous carcinoma exhibit wide inter- and intra-tumoral heterogeneity from a genomic and molecular point of view, resulting from different spatial distribution of “subclones” and their different mutations and gene expression over time [34]. This statement is equally true in residual lesions after treatment, in relation to chemoresistance acquired after therapy [35]. Radiomics with machine-learning (ML) and deep-learning (DL) algorithms makes possible the assessment of different tissue microarchitectures, processing quantitative data on small tissue volumes and distinguishing different “microenvironments” and different molecular profiles in the lesion itself and in the different disease localizations [36,37,38,39]. Non-invasive analysis of tumour heterogeneity opens the door to the so-called “virtual biopsy”, which, in association with genomic, pathologic, clinical and laboratory variables, has the potential to improve the study of neoplasm evolution and identify the best therapeutic approach for patients.

The main areas in which AI is influencing ovarian cancer imaging include:

• Classification of lesions: although histopathological examination after biopsy is still considered the gold standard for discriminating benign from malignant lesions, deep learning algorithms (a subcategory of machine learning) can classify ovarian lesions as either benign or malignant with greater accuracy than traditional techniques used by radiologists. Li et al [40] described a radiomics model capable of differentiating malignant and benign ovarian lesions in CT images of about 143 patients. Good diagnostic accuracy has also been found in radiomics applied to MRI imaging, as described by Saida et al [41] and Wang et al [38]. In the latter study, DL algorithms were able to differentiate borderline ovarian lesions from malignant ones with higher accuracy than radiologists. Also in the field of ultrasonography recent studies have demonstrated levels of diagnostic accuracy by DL algorithms comparable to O-RADS-US and experienced sonographers [42,43].
• Prediction of genetic alterations: as already well known, in patients diagnosed with ovarian cancer it is critical to establish whether the BRCA-1 and BRCA-2 genes are mutated, since BRCA-mutated tumors are associated with increased chemosensitivity to platinum-based drugs resulting in increased PFS. [44] Despite Meier et al. found no significant correlation between radiomic features and BRCA mutational status [45], other authors were able to predict Ki-67 status by analyzing radiomic features derived from PET-CT images [46].
• Prediction of disease spread at diagnosis: AI can automate the process of image segmentation, i.e. the isolation and analysis of suspicious areas in medical images. This segmentation allows radiologists to focus more accurately and quickly on potentially pathological areas. In fact, although CE-CT is the gold standard for staging ovarian cancer, it has accuracy limitations in identifying small peritoneal implants (< 1 cm) and localizations in specific areas such as the small bowel and mesentery. In addition, lesions are often “unmeasurable” according to RECIST 1.1 criteria [47]. Several studies describe how AI can predict the presence of peritoneal carcinosis and lymph node metastasis in HGSOCs on both CT and MR imaging by integrating radiomics with both clinical and laboratory factors such as age and CA-125 blood-levels [48,49,50].
• Prediction of treatment response: the prediction of treatment response according to radiomics models deeply traces the analysis of the previously described intrinsic heterogeneity of the tumor and various microenvironments. It is evident that different “subclones” of tumor tissue exhibit varied responses to different drugs in relation to their histological and molecular features. Indeed, it has been shown that while the number of disease localizations at diagnosis correlates significantly with treatment response, there is no correlation between disease volume and therapy response [51]. Conversely, when combined with clinical and laboratory data, radiomic biomarkers have been found to accurately predict response to neoadjuvant chemotherapy (NACT) [51,52]. Similarly, some studies based on ML and DL algorithms seem to be able to predict the probability of platinum-resistance of high-grade serous carcinoma [36,53]. Consequently, even in post-NACT imaging re-evaluations, any residual tumor and/or new disease localization can be accurately characterized by describing the microarchitecture and estimating the “subclone” of origin. The same applies to patients who are candidates for immunotherapy, in whom reduced intratumoral heterogeneity seems to be associated with better response [54].
• Prediction of Risk of Recurrence: Several studies have focused on estimating progression-free survival (PFS) and overall survival (OS) in order to identify patients at higher risk of recurrence [55]. Some studies have demonstrated a significant relationship between radiomic features in CT imaging and OS, specifically showing that lower tumor heterogeneity values are associated with higher OS [45,56]. Similarly, other studies have examined the relationship between CT radiomic features of ovarian masses [57,58] and peritoneal implants [56] and PFS, which has also been found to correlate with tissue heterogeneity. Rizzo et al. [57] demonstrated that three radiomic variables—specifically, the gray level run length matrix (GLRLM), 3D morphological features, and the gray level co-occurrence matrix (GLCM)—are significantly associated with disease progression at 12 months. According to Zagari et al., among the variables considered in their study that significantly correlated with PFS, shape and density appeared to have the strongest correlation [59]. Furthermore, high tumor tissue heterogeneity values have been associated with a higher risk of incomplete surgical resection (R≠0) in non-BRCA-mutated patients [45]. Conversely, Vargas et al. reported that lower heterogeneity values were associated with greater surgical resectability [56] and, consequently, with higher OS and PFS values [5]. The integration of PET radiomic features with CT features, in addition to clinical variables, further improves prognostic accuracy compared to models based solely on CT imaging [60]. Even in the context of MRI imaging, a nomogram based on radiomic features derived from MRI images, combined with clinical variables, has shown a good ability to identify patients at risk of disease recurrence [61].
• Integration with Other Data Sources: AI can combine information from different diagnostic modalities (e.g., imaging, clinical history and laboratory results) to provide a more comprehensive and precise overview of the patient’s condition. As previously mentioned, most of the studies cited have incorporated alongside imaging features clinical variables such as age, FIGO stage, serum CA-125 levels, and the presence of residual tumor. The integration of such data into ML and DL algorithms is essential for obtaining more accurate information and achieving higher levels of diagnostic precision (Figure 7).

The primary limitation of AI-based assessments is the lack of standardization and reproducibility [33]. Currently, different AI models are being used with various software and without standardized acquisition protocols. Radiomic variables vary significantly due to factors such as imaging quality, tumor location, and the technology employed; therefore, AI must be capable of handling this variability to be effective in diverse clinical settings.

Additionally, the "black box" nature of certain neural networks can make it challenging to interpret how the algorithm arrives at a specific conclusion. Other limiting factors include costs, processing time, and the current need for manual adjustments, which hinder the AI integration into clinical practice. Consequently, despite advancements, the adoption of AI in ovarian cancer imaging still requires clinical validation and close supervision by healthcare professionals. AI should be used as a decision-support tool rather than a replacement for human clinical judgment.

4. Conclusions

Ovarian cancer is one of the most common and severe gynecologic malignancies, with high global prevalence and mortality rates. Imaging in ovarian cancer is rapidly evolving with the integration of advanced technologies such as artificial intelligence. The potential of AI in medical image analysis provides significant advantages in improving early diagnosis, staging, and treatment personalization. However, further research and development are needed to overcome current challenges and ensure the safe and effective adoption of AI in clinical practice.