JillSunday/Breast-Tumor-Prediction-Model-in-African-Healthcare-Setting

This repository contains code used to develop a breast tumor segmentation and prediction model, created using open source breast ultrasound dataset, and out-of-distribution data created via data augmentation # Breast Tumor Segmentation in Low-Resource African Healthcare Settings 🏥 (Accepted for presentation at the Deep Learning Indaba 2024, Dakar, Senegal) This project addresses the challenge of breast tumor segmentation in African healthcare environments, where limited access to high-quality imaging devices and datasets hinders robust model development. Using a curated ultrasound dataset and synthetic out-of-distribution (OOD) data augmentation, we demonstrate the efficacy of nnU-Net for tumor segmentation under resource-constrained conditions. ## Table of Contents - Dataset - Project Structure - Methodology - Data Preparation - OOD Data Generation - Model Training - Results - Predictions - References - Notebooks - License --- ## Dataset ### Original Dataset The project utilizes the **Breast-Lesions-USG** dataset from The Cancer Imaging Archive (TCIA): **Citation**: Pawłowska, A., et al. (2024). A Curated Benchmark Dataset for Ultrasound Based Breast Lesion Analysis (Breast-Lesions-USG) (Version 1) [dataset]. The Cancer Imaging Archive. DOI: 10.7937/9WKK-Q141. ### Curated Dataset To simplify multi-tumor annotations, we removed auxiliary masks (e.g., `casenumber_other1.png`). The curated dataset is available on Google Drive. The synthetic OOD generation code is located in the 'OOD Data' folder --- ## Project Structure Breast-Tumor-Prediction-Model-in-African-Healthcare-Setting/ │ ├── Data/ # Raw and curated datasets ├── Data Preparation/ # Scripts for Resizing and folder arrangment of data ├── OOD Data/ # Synthetic African-like ultrasound images ├── Data Merger/ # Code for merging OOD and original data ├── nnU-Net/ # nnU-Net configuration files ├── Training/ # Training scripts ├── Predictions/ # Model predictions on test data ├── Results/ # Visualizations and metrics ├── requirements.txt # Python dependencies ├── LICENSE # MIT License └── .gitignore # Excluded files (data, logs, etc.) --- ## Methodology ### Data Preparation To address multi-tumor annotations, w …