3D Convolutional Neural Networks for MRI Brain Classification

The code was written by Natasha Basimova, Nikita Mokrov, Ilya Selnitskiy and Ilya Zakharkin.

Prerequisites

• Linux or macOS
• Python 3
• CPU or NVIDIA GPU + CUDA

Getting Started

Get Access and Download data:

We tested the performance of the proposed networks on the data from Alzheimer’s Disease Neuroimaging Initiative (ADNI) project that provides a dataset of structural MRI scans. For dowload data you need to get access. Then go to dataset. Filter all data with this parameters:

• Weighting: 1T
• Acquision Type: 3D
• Filed Strength: 3 Tesla
• Slice Thickness: 1mm

We use this dataset to test our models’ performance for a task of classifying MRI scans of subjects with Alzheimers disease (AD), early and late mild cognitive impairment (EMCI and LMCI), and normal cohort (CN).

Installation

• Clone this repo:

git clone https://github.com/izaharkin/mri-alzheimer
cd mri-alzheimer

• Install python 3.6 and all necessary requirements:
  • For pip users, please type the command pip install -r requirements.txt.
  • Also you should install fsl-lib for skull cutting

Train/test models

• For getting good perfomance of model, prepocess data by cutting skull and run:

python3 brainiac/data_preprocessing/brain_extraction.py

And also you should run two notebooks: Data Processing and Process Cut Data.

• Train classification model:

#!./scripts/train_cyclegan.sh
python3 train.py --model ResNet152

• All logs and the pretrained model are saved at unique folder./trained_model/{model}/{other parameters} as log{number}.log and model_epoch{number}.log. To view training results and loss plots, run notebook LogParser

• For getting all changeble parameters run:

python3 train.py --help

For example:

• Standart train parameters: --num_epoch 200 --batch_size 4 --optimizer Adam --lr 3e-5 --weight_decay 1e-3
• Use augmntation (random rotation and noise): --use_augmentation True
• Use sampling (oversampling and undersampling): --use_sampling True --sampling_type over
• Apply a pre-trained model: --use_pretrain True --path_pretrain PATH

Visualization

For visualization we use popular methond GradCAM with implementation. You can find it in notebook Visualisation GradCAM