Diagnostic Accuracy of Ovarian Cancer Using Convolutional Neural Network: A Systematic Review and Meta-Analysis Publisher Pubmed



Allahqoli L ; Karimzadeh A ; Kazemi Abadi A ; Aghamohammadi S Z ; Hakimi S ; Rahmani A ; Fallahi A ; Salehiniya H ; Lagana A S ; Bhagavathula A S ; Ghiyasvand M
Source: BMC Medical Informatics and Decision Making Published:2026

Abstract

Background: Accurate detection of ovarian cancer is crucial for effective treatment and patient survival. Objectives: This study aims to evaluate the diagnostic performance of convolutional neural network (CNN) algorithms for the identification of ovarian cancer. Search strategy: In this systematic review with meta-analysis, we conducted a comprehensive literature search across four electronic databases: Medline (PubMed), Scopus, Embase, and Web of Science (WOS) in June 2024 and was subsequently updated on 1 February 2026. The search strategy was developed in consultation with domain experts and information specialists to maximize both sensitivity (SE) and specificity (SP). A combination of Medical Subject Headings (MeSH) and free-text terms related to “ovarian cancer,” “convolutional neural networks,” “deep learning,” and “artificial intelligence” was used, with Boolean operators (“AND,” “OR”) applied to combine search terms effectively. Selection criteria: Our review included all observational studies evaluating CNN algorithms for ovarian cancer detection, regardless of geographical location. Study selection was managed using EndNote and involved a two-step screening process, with titles/abstracts and full texts independently assessed by reviewers. Studies reporting the diagnostic performance of CNN algorithms for histopathologically confirmed ovarian cancer were eligible for inclusion. For the meta-analysis, we included studies that provided extractable data on true positives, false positives, true negatives, and false negatives, or threshold-specific SE and SP that could be converted into a 2 × 2 format. Data collection and analysis: Data were analyzed using R (version 4.2.3). Pooled SE, SP, and Area Under the Curve (AUC) were calculated using a multilevel hierarchical model with a study-level random effect. Four subgroups—imaging modalities, CNN architectures, learning algorithms, and database types—were investigated. Meta-regression was performed, and potential publication bias was assessed using Deeks’ funnel plot of log(DOR) versus 1/Effective Sample Size. Results: Following a review of 1,043 publications on CNN algorithms for ovarian cancer detection, 47 studies were included in the systematic review and 20 in the meta-analysis. Pooled analysis showed that CNN algorithms achieved a SE of 0.94 (95% CI 0.92–0.96), SP of 0.95 (95% CI 0.90–0.97), and an AUC of 0.974 (95% CI 0.961–0.981). Among imaging modalities, magnetic resonance imaging (MRI) demonstrated the highest diagnostic performance (SE 0.97, SP 0.955, AUC 0.986), followed by computed tomography (CT) (SE 0.914, SP 0.975, AUC 0.983), histopathology (SE 0.979, SP 0.934, AUC 0.981), and ultrasound (SE 0.891, SP 0.951, AUC 0.922). Among CNN architectures, other architectures achieved the highest pooled AUC (0.979), followed by ResNet (SE 0.92, SP 0.947, AUC 0.969) and DenseNet (SE 0.927, SP 0.932, AUC 0.956). Transfer learning (SE 0.942, SP 0.948, AUC 0.978) outperformed fully trained models (SE 0.959, SP 0.929, AUC 0.962). Open-source datasets showed higher performance (SE 0.98, SP 0.971, AUC 0.985) than non-open datasets (SE 0.931, SP 0.938, AUC 0.968). Meta-regression indicated that the “other” algorithm family was significantly associated with higher logDOR, while imaging modality, dataset openness, transfer learning, and DenseNet were not significant predictors. Substantial heterogeneity remained across studies, but leave-one-out analysis confirmed the robustness of the pooled estimates, and Deeks’ test suggested potential publication bias. Conclusion: CNN-based algorithms demonstrate high diagnostic accuracy for ovarian cancer detection, with particularly strong performance across imaging modalities such as MRI, CT, and histopathology. These findings highlight the potential of deep learning models to support AI-assisted diagnostic workflows and improve early detection. However, substantial heterogeneity across studies and potential publication bias indicate the need for standardized imaging protocols, larger multi-center datasets, and external validation. Future research should focus on harmonizing data sources and integrating CNN-based tools into clinical decision-making to enhance diagnostic reliability and patient outcomes. © The Author(s) 2026.