Abstract
Overview This repository contains the source code and selected experimental results for the study Deep Learning-Based Pediatric Colorectal Polyp Classification. It provides implementations for training deep learning models, evaluating different training strategies, analyzing duplicate and near-duplicate images, conducting sensitivity analyses, performing cross-dataset evaluation, and generating Grad-CAM-based explainability visualizations. Contents Model training: Training scripts for ConvNeXt-Small, ConvNeXt-Tiny, ResNet-50, DenseNet-121, and EfficientNet-B0.Ablation studies: Scripts for comparing training strategies, including training from scratch, single-phase training, and two-phase training.Duplicate analysis: Scripts for identifying exact and near-duplicate images.Sensitivity analysis: Evaluation of model performance after excluding identified duplicate and near-duplicate images.External evaluation: Scripts for evaluating trained models on the Kvasir and VIM-Polyp datasets.Explainability: Grad-CAM and Grad-CAM++ visualizations.Experimental results: Selected CSV files and figures corresponding to the reported analyses.Reproducibility: Configuration details, random seeds, and instructions for running the experiments. Datasets and execution The CP-CHILD datasets are required for the primary experiments. Kvasir and VIM-Polyp data are required for the corresponding external evaluation experiments. The datasets are not included in this repository; users must obtain them from their respective sources and arrange the files according to the project instructions. The repository includes training and evaluation scripts. Users can install the required Python packages, configure the dataset paths, and run the scripts to train models and reproduce the corresponding analyses. Trained model checkpoints are not included unless explicitly provided. Intended use This repository is intended for research and reproducibility in deep learning-based pediatric colorectal polyp classification. The experimental results are intended for research evaluation and should not be interpreted as evidence of clinical readiness.