Breast and cervical cancers are significant global health issues, disproportionately affecting women in low- and middle-income countries (LMICs) due to late-stage diagnoses and limited access to early detection. Early detection is crucial for improved prognosis and survival. Medical imaging (mammography, ultrasound, cytology, colposcopy) plays a vital role, but its effectiveness is hampered by fragmented healthcare systems in LMICs, requiring skilled personnel and potentially leading to inter-rater variability. Deep learning (DL), a subset of artificial intelligence (AI), offers a promising solution for automating medical image analysis and improving cancer detection. While some DL-based diagnostic tools have received FDA approval, further critical appraisal is necessary. This study aims to systematically review and meta-analyze the diagnostic performance of DL algorithms in detecting breast and cervical cancers using medical imaging, addressing the need for a comprehensive assessment of this technology's potential.

The authors reviewed existing literature on the use of deep learning in medical imaging for cancer detection, noting a lack of comprehensive, medical imaging-specific systematic reviews, particularly focusing on breast and cervical cancers. They mention several studies with varying scopes and methodologies, highlighting inconsistencies in reporting and the need for more rigorous research. Existing reviews covered diverse domains, limited in scope to breast and dermatological cancers, or suffered from high heterogeneity due to methodological variations. The authors point to the need for expanded research in this area, improved standardization of methods, and improved reporting guidelines to ensure the reliability and generalizability of deep learning models.

This systematic review and meta-analysis followed the PRISMA guidelines and was registered with PROSPERO. Databases (Medline, Embase, IEEE, Cochrane Library) were searched until April 2021, with broad search terms and no language or region restrictions. After screening titles and abstracts, 71 full-text articles were assessed for eligibility, resulting in 35 studies included in the qualitative synthesis and 20 in the meta-analysis. Data extraction focused on diagnostic performance (sensitivity, specificity, AUC) from contingency tables. Hierarchical summary receiver operating characteristic (HSROC) curves were used to analyze the pooled performance of DL algorithms. Subgroup analyses were conducted based on validation type (internal vs. external), cancer type (breast vs. cervical), imaging modality, and comparison to human clinicians. The QUADAS-2 tool assessed the quality of included studies. Statistical analyses, including meta-regression and heterogeneity assessment, were performed using STATA and R.

The meta-analysis of 20 studies yielded a pooled sensitivity of 88% (95% CI 85–90%), specificity of 84% (95% CI 79–87%), and AUC of 0.92 (95% CI 0.90–0.94) for DL algorithms in detecting breast and cervical cancers. Subgroup analyses revealed comparable performance across various categories: * **Validation type:** Internal validation showed slightly higher sensitivity (89%) than external validation (83%), highlighting the potential for overestimation in internally validated studies. * **Cancer type:** Breast cancer detection showed slightly higher pooled sensitivity (90%) and specificity (86%) compared to cervical cancer detection (83% sensitivity, 80% specificity). * **Imaging modality:** Ultrasound showed the highest sensitivity (91%), while colposcopy had the lowest (78%). Mammography and cytology showed intermediate results. * **DL algorithms vs. clinicians:** DL algorithms exhibited comparable performance to human clinicians (pooled sensitivity and specificity around 87% and 83% respectively for both), with similar AUC values (0.92 for both). High heterogeneity was observed across all analyses (I² >95%), suggesting significant variation in study methodologies and potentially influencing the results. Meta-regression analysis confirmed the impact of various covariates on diagnostic performance. Publication bias was not detected, but the dominance of retrospective studies and the potential for reporting bias were acknowledged.

The findings suggest that DL algorithms may offer comparable diagnostic performance to human clinicians in early breast and cervical cancer detection using medical imaging. However, the substantial heterogeneity and methodological limitations in the included studies warrant caution in interpreting these results. The over-reliance on internally validated studies, variations in imaging modalities, and the lack of standardized reporting are significant concerns. The need for more rigorous, prospective studies with standardized methodologies is emphasized, particularly studies with external validation and clear reporting of methods. Addressing these limitations is crucial for accurately assessing the true potential and clinical applicability of DL algorithms in cancer diagnosis. The study also highlights the need for a shift from a simple “DL vs. clinicians” paradigm to a more integrated “DL+clinicians” approach for optimal workflow integration.

Deep learning algorithms demonstrate potentially comparable performance to human clinicians in early breast and cervical cancer detection via medical imaging, based on available data. However, significant methodological limitations and heterogeneity necessitate caution. Future research should focus on rigorous, standardized, prospective studies with external validation and transparent reporting to enable accurate assessment and clinical integration of these technologies. Development of disease-specific guidelines and collaborative efforts between computer scientists and clinicians are crucial for advancing this field.

The main limitations include the high heterogeneity across studies, the predominance of retrospective studies (with limited external validation), variations in study methodologies and reporting quality, and potential reporting bias favoring positive results. The relatively small number of studies for some subgroup analyses (e.g., specific imaging modalities) also limits the generalizability of those findings. The focus on histopathology as the reference standard may have excluded studies with promising results but lacking confirmatory histopathological data. The black box nature of some DL models and concerns regarding model generalizability across diverse populations are additional considerations.