Dual-branch hybrid-feature learning for imbalance-aware paediatric colonoscopy decision support with interpretable multi-scale networks

Colorectal neoplasia, while rare, is high-risk in children. As such, colonoscopic detection of paediatric polyps is of clinical importance. Current deep learning–based computer-aided systems are trained on adult data and do not provide solutions for class imbalance, cross-centre generalisation or interpretability in paediatrics. We introduce an imbalance-aware dual-branch hybrid-features (IDHF) framework for polyp versus non-polyp classification. IDHF, based on a ResNet-۵۰ backbone, is equipped with complementary texture branch to augment deep semantic features, adaptive gating for late fusion and optimisation of a class-weighted loss function with false-negative penalty and branch-agreement regularisation. The proposed model is trained on CP-CHILD-A (۸,۰۰۰ images, ۱,۰۰۰ polyp/۷,۰۰۰ non-polyp) and tested without fine-tuning on CP-CHILD-B (۱,۵۰۰ images, ۴۰۰ polyp/۱,۱۰۰ non-polyp) on another platform. ResNet-۵۰+IDHF achieves ۹۹.۶% accuracy, ۱۰۰.۰% sensitivity, ۹۹.۵% specificity, ۹۶.۸% precision and a ۹۸.۴% F۱-score on CP-CHILD-A. ResNet-۵۰+IDHF achieves ۹۹.۵% accuracy, ۹۹.۲% sensitivity, ۹۹.۶% specificity, ۹۹.۰% precision and a ۹۹.۱% F۱-score on CP-CHILD-B, supporting a robust and interpretable solution for computer-aided paediatric polyp detection.

کلیدواژه ها:

Colonoscopy ، Colorectal Polyps ، Deep Learning ، Class Imbalance ، Interpretability