Beginner's guide to MRI brain scans classification using Machine learn

Beginner's guide to MRI brain scans classification using Machine learning

START

Prerequisites

Install Docker (with GPU support if possible)

Download original IXI dataset

Clone tutorial repo from github

Obtain Freesurfer licence

Download preprocessed IXI dataset

https://shorturl.at/ilvX7

Preprocessed IXI data on dropbox

https://brain-development.org/ixi-dataset/

Original IXI data. Download demographic spreadsheet, and optionally, T1 images if you want to do preprocessing by yourself.

https://github.com/damianpolakpjwstk/dsc-2023-mri-classification

git clone https://github.com/damianpolakpjwstk/dsc-2023-mri-classification.git && cd dsc-2023-mri-classification && git submodule init && git submodule update

https://surfer.nmr.mgh.harvard.edu/registration.html

Get the FreeSurfer license

https://docs.docker.com/engine/install/

Install Docker (with GPU support if possible)

IXI dataset

https://brain-development.org/ixi-dataset/

• 600 MR images from normal, healthy subjects
• T1, T2 and PD-weighted images (we'll be using T1 weighted)
• Demographic informations (age, sex, height, weight, etc.)
• Publicly accesible

MRI data characteristics

1. 3D arrays
2. Pixels vs voxels
3. Intensity value
4. Slices and axes
5. Sequences

source: https://www.hindawi.com/journals/cmmm/2015/450341/

Axial, coronal and sagittal axis

source: https://users.fmrib.ox.ac.uk/~stuart/thesis/chapter_3/image3_5.gif, https://www.researchgate.net/figure/MRI-planes-for-MRI-head-scan-a-Axial-b-Coronal-c-Sagittal-MR-scanner-can-generate_fig2_338448026

source: https://www.researchgate.net/figure/MRI-pulse-sequences-usually-used-in-a-clinical-setting-T1-w-provides-an-anatomical_fig1_331209311

Two ways to work with MRI scans using ML

Tabular data extraction vs MRI scans as 3D images

1. Extract volumetric parameters of each structure using specialized software, eg. FreeSurfer or FastSurfer
2. Perform classification of tabular data using classical ML approach

1. Do all necessary preprocessing steps (intensity normalization, skull stripping, registration, etc)
2. Perform classification of 3D data using 3D ConvNets (eg. Pseudo-3D models)

VS

From MRI scan to tabular data

Acquiring volumetric data from scans using FreeSurfer

This process takes hours for a single scan!

Raw MRI scans in DICOM format

Tabular data - volumes of specific anatomical regions in the brain

source: https://www.imaios.com/en/e-anatomy/brain/mri-brain

From MRI scan to tabular data

Acquiring volumetric data from scans using FastSurfer

Using FastSurfer (deep learning based solution), we can extract volumetric data from brain scan in less than a minute!

source: https://deep-mi.org/research/fastsurfer/

VOLUMETRIC DATA EXTRACTIONFASTSURFER PIPELINE

1. Run: docker build -t scans . 2. Change paths in run_fastsurfer.py file 3. Run: python run_fastsurfer.py

TABULAR DATA CLASSIFICATION/REGRESSION

SCANS preparation steps

3D ConvNets

Memory and time consuming approach

source: https://medium.com/convolutional-lstm-neural-networks-for-video-frame/neural-networks-for-classification-of-spatio-temporal-data-beb1693056f

Pseudo-3D CNN models

Learning Spatio-Temporal Representation with Pseudo-3D Residual Networks

• IDEA: Instead of performing one memory-consuming 3D convolution, do two 2D convolutions
• Developed for video classification, can be also used for MRI applications

source: https://arxiv.org/abs/1711.10305

3D Convnets training

Other applications

Ideas for your own projects

Dementia progression/Alzheimer disease detection

Explainability of CNN models (GradCAM, saliency maps)

Tumor/cancer segmentation (not only in brains)

Using Pseudo-3D network to classify videos

Mental disorders detection

THANK you!

polakd7@icloud.com