Deep learning-based cervical cancer T-staging using MRI: multi-structure segmentation and classification

AIMS: Cervical cancer has high incidence and mortality, seriously threatening women's survival and quality of life. Radiologists currently rely mainly on subjective clinical experience for cervical cancer T-staging, which easily leads to misdiagnosis. OBJECTIVES: To develop a deep learning-based technique for automatic segmentation and T-staging of cervical cancer to improve clinical diagnostic accuracy and efficiency. MATERIALS AND METHODS: A dataset of 17,479 fT1WI MRI scans from 144 patients (T1-T4 stages) was constructed; tumors and adjacent structures were manually outlined with Amira 2019 to generate ground truth (GT). A novel segmentation network (CPANet) was designed by integrating global pyramid guidance (GPG) and atrous spatial pyramid pooling (ASPP) modules into CNN. CPANet's segmentation performance was validated against GT and compared with UNet, UNet++, DeepLabv3+, and UperNet-Swin. Based on CPANet-extracted ROIs, T-staging models were built using ResNet50, DenseNet121, and Swin Transformer (pathological T-staging as GT), and the optimal model was selected. RESULTS: CPANet outperformed other networks, with Dice similarity coefficients (DSCs) of 0.783 (tumor), 0.901 (uterus), 0.909 (bladder), and 0.892 (rectum), and average per-case processing time of 1.60 s. Swin Transformer achieved the best T-staging performance: AUCs of 0.713 (T1), 0.799 (T2), 0.845 (T3-T4) for main stages, and 0.623 (T1b), 0.673 (T2a), 0.897 (T2b) for sub-stages from MRI images. This not only improves diagnostic accuracy and efficiency but also conserves medical resources, thereby facilitating the establishment of intelligent healthcare systems in medically underserved areas. CONCLUSIONS: CPANet and Swin Transformer enable accurate automatic segmentation and T-staging of cervical cancer from MRI images, improving diagnostic accuracy and efficiency, saving medical resources, and facilitating intelligent healthcare in underserved areas.