nntrongnghia/tdsi21-shoulder-muscle-segmentation

MRI shoulder muscle segmentation using nnUNet

By Clément Nicolas--Graffard, Guillaume Picaud and Ngoc Trong Nghia Nguyen of INSA Lyon - 5GE - TDSI - 2021.

This is an academic project of TDSI program (Image and Signal Processing) at INSA Lyon under the supervision of Professor Thomas Grenier, Ph.D.

The original README of nnUNet can be found here

Our interpretation of nnUNet is resumed in the Wiki pages of this repository. --> see here

Install this nnUNet package

• Remove old nnUNet
• Clone this repo
• Activate your PyTorch environment
• Install nnUNet in this repository:

cd TDSI21-Shoulder-Muscle-Segmentation
pip install -e .

New features added

RMSProp

To use RMSProp in training:

nnUNet_train MODEL_NAME rmsPropTrainer Task500_Epaule FOLD --npz

TransUNet

TransUNet for 2D images is added in this nnUNet using the code in the official repository. Then we add our 3D version by adapting 2D convolution layers to 3D.

For more detail about training, inference and evaluation, please check Wiki/TransUNet 2D & 3D

Model summary

Once the training is done, you can check the number of parameters, intermediate tensor shape and model structure by nnUNet_model_summary

nnUNet_model_summary -t TASK_NAME -m MODEL_NAME

Minor modifications

• Early stopping with patien of 50 epochs
• Elastic Deformation in data augmentation is available through --do_elastic
• Advanced metrics for evaluation: Hausdorff Distance and Average Symmetric Surface Distance