Deep Learning in Gynecologic Cancer Diagnosis: Current Advances, Challenges, and Future Directions
Keywords: Artificial intelligence; Deep Learning; Cervical cancer; Endometrial cancer; Gynecologic cancer; Ovarian cancer
Abstract: Gynecologic cancers remain a significant health challenge globally, particularly in low-resource settings where diagnostic consistency and expertise are often limited. Traditional diagnostic methods based on human image analysis can be inconsistent and prone to errors. Recent advancements in deep learning (DL) offer promising solutions for automating image analysis, providing more objective and accurate diagnostic outcomes. This review synthesizes current advances, identifies challenges, and explores future directions in the application of DL for gynecologic cancer diagnosis using various imaging modalities. A thorough literature review following PRISMA-2 guidelines examined studies employing DL to diagnose various gynecologic cancers through imaging modalities like MRI, CT scans, Pap smears, and colposcopy. Data extraction and quality assessment were conducted using the QUADAS-2 tool, and diagnostic performance was evaluated through pooled sensitivity, specificity, SROC curves, and AUC metrics using R software. From 48 studies reviewed, 24 met the inclusion criteria for the meta-analysis. The studies employed various DL models, predominantly ResNet, VGGNet, and UNet, across different imaging modalities. DL models demonstrated superior sensitivity (89.40%) but slightly lower specificity (87.6%) compared to traditional machine learning (ML) methods, which exhibited 68.1% sensitivity and 94.1% specificity. The AUC for DL models was 0.88, underscoring their high diagnostic accuracy. Challenges such as study heterogeneity and methodological biases were identified, underscoring the importance of standardized protocols. Despite these obstacles, DL holds significant promise in gynecologic cancer diagnosis, particularly in resource constrained settings. Addressing these challenges can enhance the clinical utility of DL and contribute to improved patient outcomes.
Submission Number: 11
Loading