ready2drop/DTS_dataset

Advancing Medical Image Segmentation: Morphology-Driven Learning with Diffusion Transformer

This repository contains the official implementation of the paper "Advancing Medical Image Segmentation: Morphology-Driven Learning with Diffusion Transformer".

Abstract

We introduce a novel Diffusion Transformer Segmentation (DTS) model for robust medical image segmentation. Our approach combines diffusion models and transformer architectures to leverage morphological information, enhancing segmentation accuracy across various medical imaging modalities 3. Benchmark Datasets We have uploaded our benchmark datasets to Hugging Face for easy access and reproducibility. These datasets cover a range of medical imaging modalities and anatomical regions:

1. Multi-organ Segmentation
   • BTCV (Beyond the Cranial Vault)
   • Contains 30 subjects with 3779 axial abdominal CT slices
   • 13 organ annotations
   • 18 cases for training, 12 for testing
2. Brain MRI
   • BraTS2021
   • Brain tumor segmentation dataset
3. Skin Lesion
   • ISIC2018
   • Skin lesion segmentation dataset

These datasets represent diverse challenges in medical image segmentation, allowing for thorough testing of segmentation algorithms across different anatomical structures and imaging techniques.

Installation

To install the necessary dependencies and download the datasets:

# Clone the repository
git clone https://huggingface.co/datasets/ready2drop/DTS.git
cd DTS-segmentation

# Install dependencies
pip install -r requirements.txt

# Download datasets
mkdir datasets
cd datasets

# BTCV dataset
wget https://huggingface.co/datasets/ready2drop/BTCV.zip
unzip BTCV.zip

# BraTS2021 dataset
wget https://huggingface.co/datasets/ready2drop/BraTS2021_Training_Data.tar.gz
tar -xzvf BraTS2021.tar.gz

# ISIC2018 dataset
wget https://huggingface.co/datasets/ready2drop/ISIC2018_Task1-2_Training_Input.zip
unzip ISIC2018.zip