Comparison of selected CNN and transformer models for choledocholithiasis classification and localization on MRCP

Abstract

This repository contains the source code and computational workflows accompanying the manuscript entitled: “Comparison of selected CNN and transformer models for choledocholithiasis classification and localization on MRCP” The codebase provides the computational pipeline used for the development, evaluation, comparison, statistical analysis, and explainability of deep-learning models for choledocholithiasis detection on magnetic resonance cholangiopancreatography (MRCP) images. The repository includes: Training and evaluation pipelines for CNN- and Transformer-based image classification models, including ResNet50 and Vision Transformer (ViT). Object-detection pipelines based on YOLO and RF-DETR. Model evaluation and prediction-generation scripts. Statistical analysis workflows, including bootstrap confidence intervals, paired bootstrap comparisons, McNemar’s test, and DeLong’s test. Detection metrics and evaluation utilities, including IoU, mean average precision (mAP), and recall. Attention-based explainability and visualization workflows for the Transformer-based model. Configuration files and reproducibility-related settings used throughout the computational experiments. The purpose of this software release is to support transparency, reproducibility, and independent verification of the computational analyses reported in the accompanying manuscript. The dataset and patient-level medical images are not included in this repository due to data-access, privacy, and/or institutional restrictions. Users seeking to reproduce the analyses must obtain access to the required dataset through the applicable institutional and ethical procedures. The exact software environment, model dependencies, and execution instructions are documented in the accompanying README and dependency files. Users should ensure that the specified software versions and required pretrained model resources are used when reproducing the reported results. This record represents the archived software version associated with the corresponding manuscript and should be cited when the code is reused, adapted, or referenced.