Deep learning MRI AD prediction

Introduction

We use deep learning method for Alzheimer's disease (AD) prediction based on structural T1 MRI. This method can generate a 0-1 AD score for any brain T1 MRI image which can be used as a biomarker to evaluate the likelihood of Alzheimer's disease. In reference [1], we use the method for prodromal AD diagnosis.

There can be other similar use cases classifying different neurological status (aging ref[3], Schizophrenia, etc.) with different imaging modalities (T2 MRI, PET, fMRI, etc.), preliminary studies.

Model training and inference

ad_score_training.py: training script with multiple hyper-parameters to tune
ad_score_inference.py: given a directory of preprocessed nifti images (*.nii), generate the AD score and associated class-activation-map (CAM) using pre-trained model file
```
python ad_score_inference.py <MODEL_FILE>
```

pre-trained model files:

model_ad_cn_adni_only_bl_ad_cn.hdf5:
- trained with ADNI subjects who are cognitively normal (CN) or AD at baseline
- used for MCI conversion (prodromal AD) classification in ADNI in reference [1]
model_ad_cn_adni_all.hdf5:
- trained with ADNI subjects who are CN or AD at any visits
- can be used for any data other than ADNI
  - see some open neuroimaging datasets in this repository

Data

Since data use agreement is required in order to access ADNI data, we only provide the essential and non-sensitive information in adni_subjects.csv. It's straightforward to join this dataset with the full dataset from ADNI after the users gaining data access.

Preprocessing

The input brain structural MRIs are pre-processed with standard neuroimaging softwares.

Steps:

FreeSurfer individual brain extraction and normalization

Standard FreeSurfer pre-processing, -autorecon1 is enough
In FreeSurfer SUBJ_DIR

 for f in *;do mri_convert $f/mri/brainmask.mgz ${OUTPUT_DIR}/${f}_mri_brainmask.nii.gz;done

FSL registration

In ${OUTPUT_DIR}

 for f in *brainmask.nii.gz;do if [ -a ${f%.nii.gz}_mni152brain_affine_tl.nii.gz ];then echo $f;else flirt -in $f -ref /usr/local/fsl/data/standard/MNI152_T1_1mm_brain.nii.gz -out ${f%.nii.gz}_mni152brain_affine_tl.nii.gz -init ${f%.nii.gz}_mni152brain_affine.mat -dof 12 -applyxfm -interp trilinear;echo $f;fi;done

Localization/Explainability

Another important piece of this series of studies is the focus on localization/explainability.

Class activation map is one way to pinpoint the regions important to the prediction.
Another method is to mask the input with some prior (brain_region_masks) and examine the performance.

References:

Feng, Xinyang, Frank A. Provenzano, and Scott A. Small. "A deep learning MRI approach outperforms other biomarkers of prodromal Alzheimer’s disease." Alzheimer's Research & Therapy 14.1 (2022): 1-11.
Feng, Xinyang, Jie Yang, Zachary C. Lipton, Scott A. Small, Frank A. Provenzano, and Alzheimer’s Disease Neuroimaging Initiative. "Deep learning on MRI affirms the prominence of the hippocampal formation in Alzheimer’s disease classification." bioRxiv (2018): 456277. [bib]
Feng, Xinyang, Frank A. Provenzano, Scott A. Small, and Alzheimer’s Disease Neuroimaging Initiative. "Detecting prodromal Alzheimer’s disease with MRI through deep learning." bioRxiv (2019): 813899. [bib]
Feng, Xinyang, Zachary C. Lipton, Jie Yang, Scott A. Small, Frank A. Provenzano, Alzheimer’s Disease Neuroimaging Initiative, and Frontotemporal Lobar Degeneration Neuroimaging Initiative. "Estimating brain age based on a uniform healthy population with deep learning and structural magnetic resonance imaging." Neurobiology of aging 91 (2020): 15-25. [bib]

RamanPndy/Deep-learning-MRI-AD-prediction