segmentation

EVA-X: X-Ray Image Segmentation

Installation

1. Follow classification installation.

2. Install mmsegmentation==0.29.1 and its corresponding packages. You could install mmsegmentation==0.29.1 from souce by run:

   git clone https://github.com/open-mmlab/mmsegmentation
   git checkout 7b09967
   pip install -e .
   

3. build MSDA:

   cd ops
   sh make.sh
   

Weights Preparation

Prepare weights following X-ray image classification. Put them in ./pretrained

NOTE: You should run python tools/interpolate14to16 for EVA-X weights to convert precision from Half to Float.

Data Preparation

1. Download SIIM-ACR Pneumothorax / RSNA Pneumonia / Lung Segmentation / Shenzhen Hospital dataset.

2. We random have splited the datasets to train/test images and released our data split. You could prepare the dataset by run:

   python data_splits/prepare_siim.py
   python data_splits/prepare_rsna.py
   

   The processed datasets will be saved in ./dataset.

Training

All of the configs are saved in configs. Each experiment has been done on a single RTX-3090 GPU. Choose the config and run an experiemnt by run:

python train.py segmentation/config/eva_x/upernet_eva_x_small_siim_50ep.py

Results

All models have been finetuned on the same settings. We will upload the finetuned weights of EVA-X, evaluate finetuned models by running:

Weights are coming soon...

python test.py segmentation/config/eva_x/upernet_eva_x_small_siim_50ep.py \
       /path/to/checkpoint \
       --eval mIoU mDice

Acknowledgements

Our codes are built upon EVA-02, mmsegmentation, MGCA, segmentation_models_pytorch. Thanks for their great work!