MED-LOADER

This repository contains code for downloading, sorting, extracting (to .mha i.e. a MetaImage) and viewing CT slices from open datasets like:-

1. MICCAI 2015 Head and Neck (HaN) Segmentation Challenge

Usage

1. Install Anaconda with python3.7
2. Install git
3. Open a terminal and follow the commands
   • Clone this repository
     • git clone [email protected]:prerakmody/medloader.git
   • Install
     • cd medloader
     • conda env create --file environment.yml --name medloader (Ensure it is in UTF-8 encoding)
       • this will take time as it need to download and install all packages
   • Use
     • conda activate medloader
     • conda develop .
       • Adds medloader as a python pacakge on your local machine
4. To download/sort/extract the MICCAI 2015 dataset
   • Keep medloader.dataloader.config.VOXEL_RESO as an empty tuple =() if you dont want to resmaple
     • Time consuming step (but it is recommended that all 3D volumes have the same pixel spacing)
   • python demo/tf_han_miccai2015_loader.py
     • This shall create a ./_data/HaN_MICCAI2015 directory with raw/ and processed/ data files for each patient
     • If medloader.dataloader.config.VOXEL_RESO is kept empty, then also set the resampled flag to False in demo.tf_han_miccai2015_trainer.params.dataloader
   • Note
     • You may want to remove patient-id=0522c0125 due to its small dimensions in the z-axis
       • If medloader.dataloader.config.HaN_MICCAI2015['GRID_3D']['SIZE'] = [96,96,96]
         • Remove it from _data/HaN_MICCAI2015/processed/train/data_3D/{*.csv}
5. To train a model
   • Run python demo/tf_han_miccai2015_trainer.py
     • You can change the params within that file if need be
     • If you have a GPU with mem < 6GB, then change the demo.tf_han_miccai2015_trainer.params.model.name key to config.MODEL_UNET3DSHALLOW
       • If you still have memory issues, then change other parameters such as demo.tf_han_miccai2015_trainer.params.model.{batch_size,prefetch_batch,parallel_calls}
   • Note:
     • To enable demo.tf_han_miccai2015_trainer.params.model.profiler
       • To run training script without sudo: Instrutions by Nvidia

CleanUp

1. Remove conda env
   • conda env remove -n medloader

TODO

• Conbvert tensorflow code to PyTorch