Ovarian cancer is the deadliest form of cancer affecting the female reproductive system. Its early clinical symptoms are often subtle, leading to diagnosis at advanced stages and resulting in poor prognosis (1, 2). It is the leading cause of mortality among gynecological cancers, surpassing cervical and endometrial cancers in terms of lethality, with a 10-year survival rate of only 35% across all stages (3). The insidious nature of ovarian cancer presents a significant diagnostic challenge, as subtle or non-specific symptoms often result in delayed clinical presentations. Early and accurate diagnosis is pivotal to reducing mortality, improving treatment outcomes, and minimizing unnecessary surgical procedures for patients presenting with ovarian masses.

Conventional diagnostic modalities for ovarian cancer include imaging tools such as computed tomography (CT), magnetic resonance imaging (MRI), serum biomarkers like CA125 and HE4, pathological biopsy, and ultrasound. CT and MRI are common non-invasive methods, but their ability to fully analyze tumor microenvironments is limited because human vision has natural limitations (3, 4). Moreover, variability in radiological interpretation often depends on the expertise of the operator, which adding a layer of subjectivity to diagnostic accuracy. Pathological biopsy, though definitive, is invasive and not always suitable for all patients (5, 6). Serum biomarkers, despite being widely used, often lack specificity, as elevated levels can occur in conditions unrelated to ovarian malignancies (7). Ultrasound, particularly transvaginal ultrasound, has emerged as one of the most accessible and cost-effective tools for ovarian tumor evaluation (5). However, conventional ultrasound diagnostics primarily rely on morphological imaging and visual assessment, which are significantly influenced by operator-dependent variability. Quantitative data embedded within ultrasound images, which reportedly carry predictive advantages over traditional imaging metrics, often remain underexplored in clinical settings (1, 8). Thus, while these diagnostic methods offer valuable insights, their limitations underscore the urgent need for innovative solutions to improve diagnostic precision and reliability.

Artificial intelligence (AI) has demonstrated remarkable potential in improving diagnostic performance, especially in the analysis of medical imaging such as ultrasound. By leveraging machine learning and deep learning technologies, AI can extract complex patterns from imaging data and provide quantitative assessments of radiographic features that are often imperceptible to the human eye (4, 9). Recent studies have demonstrated that AI-assisted diagnostic performance in ovarian cancer can achieve sensitivity and specificity rates as high as 81% and 92%, which outperforming traditional imaging-based diagnostics (1). AI-based models can enhance the interpretation of ultrasound images by systematically quantifying tumor characteristics such as lesion size, echotexture, and morphological irregularities (8). However, despite their promising potential, AI based on ultrasound applications face significant controversies. First, the generalizability of AI models needs further evaluation, particularly regarding their diagnostic performance in external validation sets (6). Second, the relative diagnostic performance of AI compared to sonographers also requires assessment (10). These challenges underscore the necessity for systematic evaluations to clarify the performance variability of AI-enabled ultrasound diagnostics for ovarian cancer.

Therefore, this meta-analysis aims to systematically assess the diagnostic performance of AI-based ultrasound models for the initial diagnosis of ovarian cancer, and compare its performance to that of sonographers.

The meta-analysis was conducted in strict accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy (PRISMA-DTA) guidelines (11).

Ultrasound-based AI demonstrated excellent diagnostic performance in the initial diagnosis of ovarian cancer, achieving AUC values of 0.98 and 0.91 in the internal and external validation sets, respectively. Additionally, our results revealed that the AUC value for sonographers was 0.87, with no overlap in the 95%CI compared to AI, indicating that the AI model significantly outperformed traditional ultrasound diagnosis. The decline in diagnostic performance in the external validation sets may be attributed to sample heterogeneity and differences in operational standards across hospitals, which could affect the model’s generalizability. In contrast, the internal validation sets was conducted in a relatively uniform environment, ensuring the model’s high performance (1, 2). The advantages of ultrasound-based AI algorithms in complex pattern recognition and data processing enable them to better identify subtle lesion features, thereby providing higher sensitivity and specificity. However, sonographers may be influenced by factors such as experience, fatigue, and subjective judgment when interpreting images, leading to lower diagnostic performance (3, 9). Therefore, AI technology in the early screening of ovarian cancer may effectively improve diagnostic accuracy and enhance patient prognosis.

Interestingly, according to the results of our subgroup analysis, the sensitivity and specificity of deep learning were 0.94 and 0.96, respectively, while those of machine learning were 0.95 and 0.94. The differences between the two were not statistically significant (both P > 0.05). This indicates that deep learning does not significantly outperform machine learning, which is consistent with Xu et al. previous findings (1). However, due to the high heterogeneity, these results should be interpreted with caution. More studies are needed in the future to confirm whether deep learning algorithms offer any additional value over machine learning algorithms in the accurate diagnosis of ovarian cancer.

We found that regarding different AI models, the sensitivity and specificity of radiomic models did not show statistically significant differences compared to the combined approach of radiomic & clinical models (both P > 0.05). This may be attributed to the inherent complexities of integrating clinical data with imaging features (34). Radiomic, which extracts quantitative features from medical images, may not fully account for the multifactorial nature of tumor behavior and clinical symptoms that influence diagnostic outcomes (35). Furthermore, the potential redundancy in the information provided by radiomic features and clinical data may not contribute to an improvement in diagnostic performance metrics (2). In the studies we included, most of the researches that combined radiomics with clinical features used machine learning algorithms, while studies that relied solely on radiomics mostly adopted deep learning algorithms. Therefore, the choice of training algorithm may also influence the current results. More research is needed in the future to determine the value of multimodal AI algorithms compared to studies using radiomics alone.

In 2022, Xu et al. (1) performed a meta-analysis evaluating the diagnostic performance of a multimodal imaging-based AI approach incorporating CT, MRI, and ultrasound for ovarian cancer. The AI-based ultrasound diagnosis demonstrated a sensitivity of 0.91, specificity of 0.87, and AUC of 0.95. Their research compared a multimodal imaging-based AI approach incorporating CT, MRI, and ultrasound with human radiologists and found that AI outperformed the radiologists (AUC: 0.93 vs. 0.85). In 2024, Mitchell et al. (3) published another meta-analysis on AI-based ultrasound diagnosis for ovarian cancer, reporting an AI sensitivity of 0.81, specificity of 0.92, and AUC of 0.87. However, only two of the included studies compared AI with sonographers, and the diagnostic performance of sonographers was not extracted in their study. In addition, their study only performed a pooled analysis of 14 articles and did not conduct in-depth subgroup analyses based on different ultrasound equipment types or various AI algorithms.

To the best of our knowledge, this the first meta-analysis to separately extract data from internal validation sets, external validation sets, and diagnostic performance data of sonographers. Compared to previous published studies, our results demonstrate higher diagnostic performance for AI models (AUC:0.98), which may be attributed to updates in new algorithms that enhance performance. Our findings also indicate that AI-based ultrasound exhibits a diagnostic performance advantage over sonographers in the initial diagnosis of ovarian cancer (AUC:0.98 vs. 0.87), which consistent with prior findings (1).

It’s worth to noted that the inclusion of studies in our meta-analysis demonstrated significant heterogeneity, potentially impacting the overall sensitivity and specificity of AI performance. We conducted a meta-regression analysis, revealing that one of the primary sources of heterogeneity was the difference between analysis (Image-based vs. Patient-based P = 0.01). Image-based studies can provide more detailed information, leading to higher diagnostic performance, which may result in overestimation of the accuracy of findings. This is largely due to their focus on specific anatomical structures and lesions, allowing for a nuanced analysis that can highlight minute details that might be overlooked in broader studies. In contrast, patient-based studies do not capture as much information, resulting in significant differences in design and methodology. However, the high heterogeneity may also be attributed to other factors: (1) For patient characteristics, the different stages of ovarian tumors and the age of patients may also influence the performance; (2) For intervention methods, the quality of ultrasound imaging itself, which depends on operator expertise and equipment, could influence AI diagnostic accuracy. These factors highlight the need for standardized reporting in future studies to facilitate more robust meta-analyses.

Our results demonstrate that AI models based on ultrasound have achieved diagnostic performance in both internal and external validation datasets, and they have the potential to become important support tools for sonographers. AI can serve as a powerful auxiliary tool, offering physicians a valuable second opinion, enhancing diagnostic reliability, and reducing the risk of missed or incorrect diagnoses. Additionally, AI has the potential to decrease unnecessary surgeries or procedures, thereby helping to lessen the clinical workload, improve diagnostic accuracy, and prevent adverse outcomes caused by missed or delayed diagnoses (32). Implementing ultrasound-based AI in primary healthcare systems could facilitate early detection and timely management of ovarian cancer. However, it is crucial to emphasize that these models should not be regarded as standalone diagnostic standards or decision-making tools but rather as supplementary aids for sonographers (9). Due to the significant differences in algorithm interpretability and the lack of standardized algorithm training for ovarian cancer diagnosis, we can only conclude that our results represent a short-term phenomenon, indicating that current AI algorithms already have the potential to outperform traditional sonographers. Notably, only five studies in our analysis compared AI diagnostic performance directly with that of sonographers, highlighting the need for further research in this area.

Some limitations of the current meta-analysis should be considered when interpreting the results. First, most of the included studies were retrospective in design, with only four employing a prospective approach, which may introduce potential biases. To assess the possible impact of different study designs on the outcomes, we conducted a meta-regression analysis, which revealed that the type of study did not significantly affect the results (both P > 0.05). Therefore, well-designed prospective studies are essential to validate the findings of this meta-analysis and ensure the reliability and generalizability of the results. Second, only the optimal AI algorithms were selected for analysis, which may lead to an overestimation of the results. This was primarily because we aimed to avoid potential patient overlap. However, this approach may have led to an overly optimistic assessment of AI model performance. Future studies should compare the performance of different models, including those with moderate or poor outcomes, to provide a more comprehensive and objective evaluation. Third, due to lack of information, the study did not stratify sonographers by their experience and skill level. Future studies should consider stratifying sonographers by experience levels to better evaluate the impact of AI assistance on diagnostic performance across different practitioner groups. Fourth, since most of the included studies classified borderline tumors as part of the malignant (positive) group, we followed their approach and classified both borderline and malignant tumors as a positive group. However, this may have contributed to an increased rate of positive diagnoses. In the future, there is a need to evaluate AI’s ability to distinguish between benign, borderline, and malignant ovarian tumors to aid early clinical intervention for malignant cases. Forth, the very high heterogeneity (It > 97%) observed in both sensitivity and specificity, partly driven by methodological factors such as the unit of analysis, significantly limits the strength of conclusions that can be drawn regarding their current clinical readiness. Therefore, any claims about their ability to enhance diagnostic reliability, reduce workload, or improve patient outcomes in clinical practice are currently premature and require validation through more standardized, large-scale studies. Fifth, the classification of borderline tumors represents a potential source of heterogeneity. Most included studies categorized borderline tumors as malignant for the purpose of binary classification. While this reflects a common clinical challenge in distinguishing these entities from invasive cancer, it may inflate the perceived sensitivity of AI models and complicate the clinical interpretation of results, as the management of borderline tumors differs from that of frank malignancy. To address this concern, we performed a sensitivity analysis excluding studies that explicitly included borderline tumors in the malignant group. Future research should aim to evaluate AI performance across a three-tiered stratification (benign, borderline, malignant) to better assess its clinical utility in differentiating this diagnostically challenging spectrum of disease.

AI based on ultrasound diagnosis demonstrates excellent performance for malignant ovarian cancer detection, with potentially superior performance compared to sonographers. Despite high heterogeneity across studies and the observed publication bias, these results indicate the potential for AI integration into clinical practice. Further studies with external, multicenter prospective head-to-head design are still needed.