MATPUSeg: Reproducibility Data, Source Code, Model Predictions and Statistical Results for Breast Ultrasound Lesion Segmentation

Description

This dataset contains reproducibility materials for MATPUSeg, a deep learning framework for breast ultrasound lesion segmentation integrating Stable Product Unit convolutions, dual spatial-channel attention, a lightweight Transformer bottleneck, and Monte Carlo Dropout uncertainty quantification. The repository contains source code, configuration files, experimental results, baseline comparisons, ablation studies, statistical analyses, uncertainty-calibration outputs, and documentation for reproducing BUSI and external-dataset experiments. Original medical images are not redistributed; acquisition instructions and dataset references are provided. MATPUSeg achieved a Dice score of 0.857, an IoU of 0.768, an AUC of 0.978, and an HD95 of 15.7 pixels on BUSI.

Steps to reproduce

Download and extract the repository and install the packages listed in source_notebook/requirements_notebook.txt. Obtain the BUSI breast-ultrasound dataset from its original authorized source and retain the benign, malignant, and normal directory structure. Use source_notebook/reproduce_notebook_split.py with random seed 42 to reproduce the deterministic dataset partition used by the uploaded notebook. Open source_notebook/jobs-breast-ultrasound-segm.ipynb, configure the local BUSI path, and execute all cells sequentially to reproduce preprocessing, training, testing, statistical analysis, and explainability experiments for PU-UNet and Residual U-Net. Reference experimental tables are provided in baseline_comparison.csv, per_class_Dice.csv, and statistical_tests.csv. The final MATPUSeg manuscript additionally uses Dual Attention Gates, a lightweight Transformer bottleneck, MC-Dropout uncertainty estimation, deep supervision, a 100-epoch training protocol, and external validation. Since the uploaded notebook represents the earlier PU-UNet/ResUNet experiment rather than the complete final MATPUSeg experiment, users should consult NOTEBOOK_PROVENANCE_AUDIT.md and MANUSCRIPT_NOTEBOOK_ALIGNMENT.csv before claiming exact reproduction of the final MATPUSeg numerical results.

Institutions

• Villa College

  Kaafu Atoll, Malé

Categories

Licence