In January 2025, a comprehensive search was conducted in three high-impact scientific databases: PubMed, IEEE Xplore, and Scopus. The search strategy included the following terms combined using Boolean operators: (“machine learning” OR “artificial intelligence” OR “deep learning” OR “neural network”) AND (“ovarian cancer” OR “ovarian tumor”) AND “ultrasound.” No language or publication type restrictions were applied during the initial search.

The following inclusion and exclusion criteria were explicitly defined to ensure transparency and reproducibility in the systematic selection of studies (see Table 1).

Studies meeting all inclusion criteria and none of the exclusion criteria were eligible for inclusion in this systematic review and meta-analysis.

Two independent reviewers (IGA and FVF) initially evaluated the title and abstract of each article identified through the database search, applying the inclusion and exclusion criteria explicitly defined in Section 2.2. This preliminary assessment allowed for the exclusion of clearly irrelevant or ineligible studies. In cases of discrepancies during this initial stage, a third reviewer (EDB) was consulted to reach consensus. Subsequently, the full texts of the preselected articles were reviewed again by both reviewers (IGA and FVF) to confirm their definitive eligibility for inclusion in the quantitative analysis.

The following variables were extracted from each study: author, year, type of segmentation, model architecture, model name, image dataset size, type of ovarian masses, number of classes, and performance metrics (accuracy, sensitivity, specificity, and AUC). For studies reporting multiple models, the one with the best overall performance was selected to avoid data duplication. The information was systematized into a structured database for subsequent statistical analysis.

The methodological quality of the included studies was assessed using the PROBAST tool (Prediction model Risk of Bias Assessment Tool) (Wolff et al., 2019), which evaluates the risk of bias and applicability in studies that develop or validate prediction models. PROBAST comprises 20 items grouped into four domains: (i) participant selection, (ii) predictors, (iii) outcomes, and (iv) statistical analysis. Two reviewers (IGA and FVF) independently performed this evaluation. Discrepancies were resolved by consensus. A detailed assessment of the 20 PROBAST items for each study and domain-specific classifications is included as Supplementary Tables S3, S4.

The analysis of the 44 included studies revealed the high average diagnostic performance of AI models applied to B-mode TVS images for OC detection. The mean accuracy was 92.3% ± 5.8, with mean sensitivity and specificity of 91.6% ± 7.2 and 90.1% ± 8.1, respectively. AUC values were reported in only 23 studies, with a mean of 0.93 ± 0.04, reflecting strong overall discriminative capacity. However, the partial availability of AUC reporting may indicate a potential reporting bias that limits the robustness of comparative analysis across all models. Beyond AUC’s limited reporting (23/44 studies), F1-score was scarcely available across the corpus. This pattern likely reflects historical reliance on accuracy/sensitivity/specificity in ultrasound AI, frequent absence of continuous model scores (hindering AUC), and the lack of confusion matrices or class-wise results needed for F1. In addition, F1 is sometimes reported as Dice in segmentation studies; because our review targets classification performance, segmentation-specific Dice metrics were not pooled, which might also contribute to the perceived under-reporting of F1.

Figure 2 provides a comparative overview of the four main performance metrics for the 10 top-performing models. This visualization highlights how specific models exhibit strong accuracy yet relatively lower specificity, an observation with important clinical implications when considering false-positive rates in diagnostic triage.

Several models, including those based on OCD-FCNN, probabilistic neural networks (PNN), and ResNet-34, reported peak performance values exceeding 95%. However, many of these models were trained and tested on small or non-external datasets, lacked proper cross-validation, or relied exclusively on internal test sets. Such methodological limitations increase the likelihood of overfitting and restrict the generalizability of reported outcomes. None of the highest-performing models reported prospective validation or integration into clinical workflows, which remains essential for evaluating real-world applicability.

To assess whether methodological design influenced diagnostic performance, non-parametric tests were conducted using accuracy as the outcome variable. A Mann–Whitney U-test revealed a statistically significant difference in accuracy between models using automatic versus manual segmentation (U = 234.0, p = 0.007), favoring automatic methods. This finding suggests that automated segmentation enhances standardization and reduces variability across studies.

Conversely, a Kruskal–Wallis test comparing performance across AI architectures (e.g., CNN, ML, ANN) did not identify statistically significant differences (H = 6.53, p = 0.258), indicating that no specific architectural family demonstrated superior accuracy within the current dataset. Nevertheless, visual inspection using violin plots (Figure 3) showed a moderately higher central tendency and reduced the variance in accuracy among CNN-based models compared to classical machine learning (ML) approaches. While this pattern may reflect the architectural strengths of CNNs in capturing spatial hierarchies within medical images (Litjens et al., 2017), it should be interpreted cautiously. CNN-based models were more frequently applied in recent studies, which may also have benefited from advances in data augmentation, automatic segmentation, and preprocessing pipelines. Therefore, the observed trend could be confounded by methodological improvements rather than an inherent advantage of architecture.

It is also important to note that model performance was evaluated exclusively using accuracy, as this was the most consistently reported metric across studies. While this allowed for comparability, it may limit interpretability in class-imbalanced settings, where metrics such as AUC or F1-score are often more informative. Future studies should prioritize the reporting of multiple complementary metrics to capture diagnostic value more comprehensively.

These descriptive findings are further explored and formally tested in the meta-regression presented in Section 3.4.

In summary, while reported performance metrics are generally high, the absence of standardized validation protocols, partial reporting of key metrics, lack of weighted or stratified analyses, and underreporting of methodological variables (especially segmentation and validation strategies) limit the interpretability and clinical generalizability of the findings. Future studies should adopt harmonized reporting guidelines (e.g., TRIPOD-AI, PROBAST-AI), employ multicenter and external validation, and report performance metrics in clinically meaningful terms to support reliable and reproducible integration into diagnostic workflows.

The relationship between the number of images used to train AI models and their diagnostic performance was evaluated using non-parametric correlation analysis. Although Pearson’s method was initially considered, the Shapiro–Wilk test confirmed that dataset size and performance metrics (accuracy, sensitivity, specificity) were not normally distributed (p < 0.001 for all), prompting the use of Spearman’s rank correlation.

Studies with more than 5,000 images were excluded from this analysis to reduce the risk of statistical distortion from highly imbalanced sample sizes. While large-scale datasets (up to 575,000 images) have become increasingly common in AI development, such volumes do not reflect typical clinical practice and may disproportionately drive correlation estimates. The 5,000-image threshold was selected to capture real-world data conditions better while preserving inter-study variability. Descriptive analysis of the full dataset showed that this threshold approximately corresponds to the 75th percentile of dataset sizes among included studies.

Spearman correlation coefficients between dataset size (≤5,000) and model performance metrics were weak and statistically non-significant. Specifically, the correlation with accuracy was ρ = 0.080 (p = 0.653), with a 95% confidence interval of −0.27 to 0.41 and an R² of 0.006, suggesting that less than 1% of the variation in accuracy could be explained by dataset size. For sensitivity, the correlation was ρ = 0.246 (p = 0.154; 95% CI: −0.09 to 0.54; R² = 0.061), and for specificity, ρ = 0.183 (p = 0.350; 95% CI: −0.20 to 0.52; R² = 0.034). Table 3 summarizes these results, including the correlation coefficients, confidence intervals, and the proportion of explained variance.

While these findings suggest that increasing dataset size within the studied range does not systematically improve model performance, this interpretation should be cautiously made. The exclusion of large datasets may limit the generalizability of these findings, and potential interaction effects, such as those involving the segmentation method, risk of bias, or model architecture, were not examined in this univariate analysis. These results are, however, consistent with the multivariable meta-regression analysis presented in Section 4, in which dataset size did not emerge as a significant independent predictor of accuracy.

One plausible explanation lies in the widespread adoption of data augmentation strategies. Techniques such as image rotation, scaling, contrast adjustment, and noise addition simulate data variability and may reduce the dependency on raw volume. However, excessive use of augmentation may also lead to redundancy or learning saturation, where additional data no longer meaningfully improves generalization.

This interpretation is in line with prior literature. For instance, Roberts et al. (2021) found that dataset size was not consistently associated with performance in a comprehensive medical imaging AI studies review. Instead, methodological rigor, validation strategy, and data diversity were identified as stronger predictors of performance. Furthermore, the risk of performance overestimation due to augmented or homogeneous datasets remains a critical concern in model evaluation.

Figure 4 presents scatterplots of accuracy, sensitivity, and specificity versus dataset size (≤5,000), each overlaid with a non-parametric LOWESS regression line and 95% confidence bands. While substantial scatters are observed across all metrics, the absence of clear or consistent directional trends underscores the importance of factors beyond sample size, such as annotation quality and experimental design in developing reliable diagnostic models.

To evaluate the impact of segmentation strategy on the diagnostic performance of AI models, performance metrics were compared between studies that implemented automatic segmentation (n = 27) and those that used manual segmentation (n = 17).

Models using automatic segmentation achieved a significantly higher average accuracy (94.2% ± 4.3) than manual segmentation (88.2% ± 6.6, p = 0.012). Sensitivity also favored automatic segmentation (93.7% ± 5.6 vs. 88.6% ± 8.3, p = 0.042). Although specificity was higher in the automatic group (92.5% ± 6.0 vs. 87.3% ± 9.5), the difference was not statistically significant (p = 0.084). AUC values were nearly identical between both groups (p = 0.839). Confidence intervals for these comparisons were not reported but are recommended for future studies to enhance the interpretability and reproducibility of statistical estimates.

Levene’s test revealed a significantly more significant variance in specificity within the manual segmentation group (p = 0.045), indicating less consistency. This is consistent with previous literature findings, highlighting manual segmentation’s susceptibility to inter- and intra-observer variability, particularly when standardized annotation protocols or multiple expert raters are not employed (Taha and Hanbury, 2015; Menze et al., 2015).

Although the primary studies reported heterogeneous segmentation details that precluded a stratified meta-analysis by architecture, a brief practical comparison is informative for clinical implementation. U-Net remains the canonical encoder–decoder with skip connections that performs well when lesion boundaries are reasonably defined, and training data are limited. AdaResU-Net augments U-Net with residual blocks and adaptive mechanisms that enlarge the effective receptive field and stabilize gradient flow, improving boundary delineation in speckle-rich ultrasound and in the presence of heterogeneous echotexture. In practice, U-Net offers simplicity and fast deployment; AdaResU-Net can yield crisper contours and fewer leakage errors near cyst walls at the cost of extra parameters. These architectural tradeoffs are likely to contribute to the higher and less variable accuracy we observed with automated vs. manual segmentation. Future primary studies should report standardized segmentation metrics (e.g., Dice, surface distance) alongside classification endpoints to enable formal architecture-level synthesis.

Figure 5 presents comparative boxplots of accuracy, sensitivity, specificity, and AUC by segmentation type. The distributions reveal higher mean values for automatic segmentation across most metrics, lower dispersion, and fewer outliers, especially for specificity and sensitivity. This visual trend suggests increased consistency, which may be attributed to the standardization benefits of automated pipelines.

Table 4 provides a detailed summary of methodological characteristics and performance metrics stratified by segmentation type. Studies using automatic segmentation not only performed better on average but also used considerably larger datasets (mean = 22,941 vs. 1,121 images) and demonstrated lower standard deviation across metrics such as accuracy (4% vs. 7%) and specificity (6% vs. 10%).

However, this interpretation should be approached with caution. Although these group-level comparisons suggest superior performance with automatic segmentation, the analysis did not control for potential confounders such as dataset size, model architecture, training methodology, or publication year. Importantly, these variables may co-vary with segmentation strategy, particularly since automatic methods are more prevalent in recent, technically advanced studies.

As shown in the meta-regression (Section 3.4), the segmentation strategy remained a significant independent predictor of accuracy even after adjusting for these covariates. Nevertheless, the unadjusted differences observed here might still reflect broader methodological convergence rather than a causal advantage of automatic segmentation per se.

Future studies should incorporate multivariable models, harmonized annotation protocols, and prospective designs to clarify the segmentation method’s isolated effect on AI model performance. Moreover, reporting standards such as STARD-AI and TRIPOD-AI should be adopted to ensure replicability and transparency in performance evaluation across studies (Collins et al., 2021; Sounderajah et al., 2021).

To further explore the drivers of diagnostic performance, a meta-regression was performed using accuracy as the dependent variable and four predictors: dataset size, segmentation strategy (automatic vs. manual), model architecture (CNN vs. other), and risk of bias (high vs. low). The regression model included 26 studies with complete data across all variables (Figure 6).

The overall model was statistically significant (F = 3.98, p = 0.015), with an adjusted R² of 0.32, indicating that the included predictors could explain approximately 32% of the variance in reported accuracy.

Among the covariates, segmentation strategy emerged as a significant predictor: studies using automatic segmentation reported on average, a 6.6 percentage point higher accuracy compared to those using manual segmentation (β = 0.0656, p = 0.007). This aligns with previous findings suggesting that automated preprocessing may reduce inter-observer variability and improve reproducibility.

Other predictors, such as dataset size, CNN architecture, and risk of bias, were not statistically significant at the conventional threshold (p > 0.050). However, the effect of high risk of bias approached significance (β = 0.0427, p = 0.096), suggesting a possible inflation of performance estimates in studies with methodological limitations.

Notably, dataset size was not a significant predictor (p = 0.323), corroborating earlier findings that performance does not linearly scale with sample size within the studied range, possibly due to saturation effects or compensatory use of data augmentation techniques.

These results reinforce the critical role of segmentation quality in shaping model performance and highlight the need for more standardized methodologies and transparent reporting in AI-based diagnostic research.

The progression of AI architectures in the included studies reflects a clear methodological shift over time. Temporal analysis of architectural usage revealed a transition from traditional ML techniques and ANNs to DL approaches, particularly CNNs. Between 2014 and 2018, studies primarily employed ML methods such as support vector machines (SVM), random forests (RF), and logistic regression (LR), representing approximately 85% of the methodologies used during this period. ANN-type architecture was also present, constituting roughly 15% of studies, while no CNN-based models were recorded before 2020.

Figure 7 presents a stacked bar chart showing the number of studies using each architecture per year. CNNs emerged in 2020, accounting for 20% of the methodologies that year, and showed a marked increase in 2021, becoming the predominant architecture (65% of studies) in 2022. This trend intensified through 2023 (72% of studies) and 2024 (78% of studies), with CNNs accounting for more than half of the models evaluated annually. Transformer-based architectures, specifically Swin Transformer and Pyramid Vision Transformer, appeared exclusively in 2024, accounting for approximately 10% of the methodologies that year, indicating the beginning of a new phase of exploration focused on models with advanced contextual attention mechanisms and long-range feature integration.

This shift mirrors broader trends observed across diagnostic imaging AI, where deep architectures have largely replaced classical ML techniques due to their ability to learn hierarchical features directly from raw images without manual feature engineering (Litjens et al., 2017; Esteva et al., 2019). However, this evolution may also explain some of the performance differences observed in earlier sections. For example, the predominance of CNNs in recent years may co-occur with advances in preprocessing, data augmentation, and training infrastructure, confounding the interpretation of architecture-based performance gains.

Notably, this trend may influence perceived model superiority, as CNN-based studies often reflect newer methodological standards, including automatic segmentation and more rigorous evaluation protocols. These temporal patterns underscore the importance of accounting for publication year and technological maturity when comparing performance across architectures or studies.

To evaluate how methodological quality influences the reported performance of AI models, the 44 included studies were classified according to their overall risk of bias using the PROBAST tool. In total, 26 studies were classified as low risk of bias, 12 as unclear risk, and six as high risk (Figure 8). However, not all studies reported all performance metrics, and the number of studies included varied substantially by metric. This heterogeneity in reporting introduces selection bias and impairs comparability.

Studies classified as high risk of bias showed markedly elevated performance metrics, with a mean accuracy of 96.1% ± 7.6, a sensitivity of 95.4% ± 7.9, and a specificity of 93.9% ± 11.7. However, none of these studies reported areas under the curve (AUC) values, precluding the complete evaluation of discriminative performance. Moreover, the elevated standard deviation in specificity suggests potential overfitting, likely arising from methodologically weak practices such as internal validation without cross-validation, small sample sizes, and operator-dependent manual segmentation (Wolff et al., 2019; Collins et al., 2021). These practices have been consistently linked to inflated model performance in machine learning for medical imaging (Kelly et al., 2019).

In contrast, low-risk studies showed more conservative but consistent performance metrics, with lower dispersion and complete reporting of AUC. Specifically, they reported an accuracy of 90.4% ± 5.2 (n = 20), a sensitivity of 90.7% ± 7.0 (n = 25), a specificity of 89.1% ± 8.2 (n = 22), and an AUC of 93.5% ± 3.9 (n = 20). While these results appear numerically inferior, the reduced variability and broader metric completeness suggest higher methodological reliability and clinical applicability.

The unclear-risk group, often the result of poor reporting rather than clearly defined methodological shortcomings, yielded intermediate metrics (accuracy: 93.6% ± 5.0; sensitivity: 91.9% ± 7.7; specificity: 91.1% ± 5.4; AUC: 92.2% ± 4.6). However, the small and variable sample sizes for each metric (e.g., n = 6 for specificity) compromise interpretability and hinder statistical power.

A sensitivity analysis was conducted excluding high-risk studies, which resulted in a notable reduction in extreme values (e.g., 100% accuracy) and decreased overall dispersion, particularly in specificity. However, this analysis was limited by the lack of formal statistical testing (e.g., ANOVA or Kruskal–Wallis) to assess whether differences between groups were statistically significant. Furthermore, no regression adjustment was made for potential confounders such as dataset size or model complexity. These omissions limit the strength of causal inferences between risk of bias and reported model performance.

The primary sources of methodological heterogeneity across studies were identified as follows: (i) reliance on manual, operator-dependent image segmentation; (ii) absence of cross-validation or external validation; (iii) small sample sizes (<200 cases); and (iv) lack of standardized metric reporting formats. These deficiencies were most prevalent in high- and unclear-risk studies, consistent with prior evidence from systematic reviews of machine learning in healthcare (Sounderajah et al., 2021; Liu et al., 2019).

Finally, this section would benefit from including confidence intervals and formal effect size estimates to contextualize differences across bias strata. Without these, claims about “superior” or “more stable” performance remain largely descriptive and potentially misleading.