CMB-Net: A Clinically Modulated Boundary-Aware Network for Anatomical Segmentation of the Cervical Transformation Zone in Colposcopy.

Colposcopy is an essential tool for cervical precancer evaluation and biopsy guidance. Most existing artificial intelligence (AI) tools for cervicogram analysis are lesion-centric or limited to global transformation zone (TZ) classification and therefore do not explicitly delineate TZ-related anatomical landmarks. Because the TZ is a major site of cervical carcinogenesis and is defined by the original squamocolumnar junction (SCJ), the new SCJ, and the external cervical os, landmark-level TZ segmentation may provide clinically meaningful support for colposcopic interpretation. To address this need, we reformulated TZ-related anatomical segmentation as a landmark-driven, four-class semantic segmentation task. We propose the Clinically Modulated Boundary-aware Network (CMB-Net), which integrates patient-specific clinical variables to account for heterogeneous SCJ visibility and incorporates multi-scale boundary supervision to improve the delineation of ambiguous anatomical interfaces. Applied to 889 internal cervicograms, CMB-Net achieved an mDice of 87.37 ± 0.17 % and an mIoU of 78.39 ± 0.24 %. In two independent external cohorts comprising 310 cases from two additional centers, it achieved an mDice/mIoU of 80.95%/69.28%, outperforming CNN- and transformer-based baselines. In the same 310-case external test cohort used for observer comparison, CMB-Net exceeded junior colposcopist-reference agreement (69.47%) and was comparable to senior colposcopist-reference agreement (80.85%). These results suggest that combining patient-conditioned priors with boundary supervision can support robust and interpretable TZ-related anatomical segmentation.