Lesion-Net: A Lesion-Oriented Hierarchical Transformer for Ischemic Stroke Segmentation in DWI

This is an official implementation of Lesion-Net: A Lesion-Oriented Hierarchical Transformer for Ischemic Stroke Segmentation in DWI

Release date: 15/Dec/2025

Abstract

Accurate segmentation of acute ischemic stroke remains challenging, particularly when lesion burden is low and abnormalities occupy only a small image region. Lesion-Net is a 2D hierarchical transformer framework that adapts spatial resolution, encoder depth, and channel capacity to this setting. It reduces early down-sampling, allocates greater depth to high-resolution stages, and maintains uniform channel widths, together with a lightweight multiscale decoder and feature-fusion head. Lesion-Net achieves DSC scores of 79.82% on ISLES 2022 and 77.14% on JHUS. Controlled ablations support the proposed stage allocation, while burden-stratified evaluation shows its strongest relative advantage in low-burden cases.

Usage

Installation

The framework was tested using Python 3.10, PyTorch 2.6, and CUDA 12.4. Ensure that you install all the dependencies listed in requirements.txt.

conda create -n lesion_net python=3.10
conda activate lesion_net
cd Lesion-Net
pip install -r requirements.txt

Datasets

The ISLES 2022 dataset is publicly available and can be downloaded from Kaggle. In contrast, the JHUS dataset is a restricted resource and can only be accessed through a formal data request submitted to ICPSR.

After placing the downloaded 3D volumes in data/isles22/3d_data/, generate 2D slices and patient-wise train/val/test splits using:

python make_dataset.py

data/
└─ isles22/
   ├─ 3d_data/ISLES-2022/
   │  └─ ...
   └─ 2d_data/
      ├─ images/
      │  ├─ train/
      │  ├─ val/
      │  └─ test/
      └─ labels/
         ├─ train/
         ├─ val/
         └─ test/

Training & Evaluation

After setting the parameters in config/config_train.yaml, run the following command to train and evaluate the model:

python train.py
python evaluate.py