Astarakee / segrap2023

Codes of model wining the SegRap MICCAI challenge 2023.

Winning submission to the SegRap 2023 MICCAI challenge

Inference models for tasks 1 (OAR segmentation) and 2 (GTV segmentation) of SegRap Challenge 2023. The submitted model won the first rank for task 2 and 6th rank for task 1 in the final(test) phase of the challenge.

This repo contains the codes and pre-trained weights for the winning submission to the SegRap 2023 MICCAI challenge. The code was developed based on nnUNet and TotalSegmentator.

Run the inference via docker

Inference codes for tasks one and two are separated into ./task1 and ./task2 subdirs.

To build the segrap2023_oar_segmentationcontainer (task1) Docker Image:

docker build . -t segrap2023_oar_segmentationcontainer

To run the prediction on new data:

docker run --rm --gpus <GPU NUM> --network none --memory="32g" -v <PATH/TO/LOCAL/INPUT_FOLDER>:/input/images/ -v </PATH/TO/LOCAL/OUTPUT_FOLDER>:/output/images/ --shm-size 2g segrap2023_oar_segmentationcontainer

Likewise, to build the segrap2023_gtv_segmentationcontainer (task2) Docker Image:

docker build . -t segrap2023_gtv_segmentationcontainer 

To run the prediction on new data:

docker run --rm --gpus <GPU NUM> -v <PATH/TO/LOCAL/INPUT_FOLDER>:/input/images/ -v </PATH/TO/LOCAL/OUTPUT_FOLDER>:/output/images/ --shm-size 2g segrap2023_gtv_segmentationcontainer

Please note that the Input Path must contain two subdirectories inclduing ./images/head-neck-ct/ (non-contrast-ct images) and images/head-neck-contrast-enhanced-ct/ (contrast-ct images) with volumetric images in .mha file format. More details regarding the data structure and formats can be found in SegRap official repo

Citation

If you found this work useful for your research, please consider citing:

arXiv{Astaraki2023, title = { Fully Automatic Segmentation of Gross Target Volume and Organs-at-Risk for Radiotherapy Planning of Nasopharyngeal Carcinoma },
author = {Astaraki, Mehdi and Bendazzoli, Simone and Toma-Dasu, Iuliana},
url = {https://doi.org/10.48550/arXiv.2310.02972}, year = {2023}}

Acknowledgment

This model was developed, evaluated, and submitted by Mehdi Astaraki, and Simone Bendazzoli.