Deep learning-based pediatric colorectal polyp classification: evaluation of training strategies, robustness, and generalization

Abstract Although automated analysis of colorectal polyps during colonoscopy has the potential to support clinical decision-making, pediatric colorectal polyp classification remains challenging because of limited annotated datasets and variability in endoscopic appearance. This study systematically evaluates a pediatric colorectal polyp classification framework employing a two-phase transfer learning strategy across multiple pretrained convolutional neural network architectures. Controlled ablation experiments were conducted on the CP-CHILD-A and CP-CHILD-B datasets using five independent random seeds. For ConvNeXt-Small, the two-phase protocol improved accuracy from $$98.88\% \pm 0.54\%$$ to $$99.77\% \pm 0.16\%$$ on CP-CHILD-A and from $$89.21\% \pm 4.23\%$$ to $$99.88\% \pm 0.27\%$$ on CP-CHILD-B compared with training from scratch. Duplicate and near-duplicate analyses followed by inference-only sensitivity evaluation showed minimal changes in performance. Cross-dataset evaluation demonstrated strong transfer between CP-CHILD-A and CP-CHILD-B, while external evaluation on Kvasir revealed substantial domain shift. Evaluation on the independent VIM-Polyp dataset showed high positive-case recall, although this dataset comprised an adult, positive-only cohort. Overall, the findings support the effectiveness of two-phase training within the evaluated CP-CHILD setting while highlighting the importance of dataset diversity, patient-level separation, and independent pediatric validation before clinical deployment.