MRI Deep Learning Projects Overview

This repository contains deep learning projects built on several large public neuroimaging datasets, covering supervised classification, regression, self-supervised pretraining, and interpretability analysis.

---

Datasets

Folder	Dataset	Summary
ABCD/	Adolescent Brain Cognitive Development	Adolescent brain development cohort with suicide/non-suicide labels and related clinical scale scores
ABIDE_I/	Autism Brain Imaging Data Exchange I	Autism spectrum disorder vs. healthy control, collected across multiple sites
ABIDE_II/	Autism Brain Imaging Data Exchange II	The second ABIDE release with a larger sample size and improved acquisition protocols
HCP_Y/	Human Connectome Project - Young Adult	High-resolution structural and functional imaging for healthy young adults, using the HCP minimal preprocessing pipeline
ADNI/	Alzheimer's Disease Neuroimaging Initiative	Longitudinal Alzheimer's cohort with five classes: CN / SMC / EMCI / LMCI / AD

All datasets are stored in preprocessed .nii.gz format, including skull stripping, bias field correction, and registration to MNI152 standard space.

---

Project Structure

MRI_Deep_Learning_Projects/
├── ABCD_Project/          # ABCD supervised learning
├── ADNI_Project/          # ADNI multi-class classification, regression, and SHAP visualization
├── SimCLR_Project/        # Cross-dataset self-supervised pretraining and feature extraction
├── ABCD/                  # Dataset directory (.nii.gz)
├── ABIDE_I/               # Dataset directory
├── ABIDE_II/              # Dataset directory
├── HCP_Y/                 # Dataset directory
└── ADNI/                  # Dataset directory

---

ABCD_Project: Suicide Risk Classification and Score Regression

Using the ABCD dataset with structural MRI as input, this project trains a 3D VGG16 model for two tasks:

Model architecture: 3D VGG16 with 5 convolution blocks, followed by AdaptiveAvgPool3d -> 4096-dim -> classification/regression head

File	Task	Label	Loss
train_onehot.py	Binary classification: suicide vs. non-suicide	label (0/1)	CrossEntropyLoss
feature_onehot.py	Extract features for the full dataset	-	-
train_float.py	Regression: risk score prediction	score (MinMax scaled to [0,1])	MSELoss
feature_float.py	Extract features for the full dataset	-	-

• Classification output: best_model_onehot.pt
• Regression output: best_model_float.pt
• Feature dimension: 4096-dim, covering all samples and saved as CSV

---

ADNI_Project: Alzheimer's Staging and SHAP Interpretability

Using the ADNI dataset, this project performs five-class cognitive staging classification and clinical score regression, then uses SHAP to locate the brain regions emphasized by the model.

Five-class label definition (see ADNI_Project/README.md for details):

Label	Meaning	Core Criteria
CN	Cognitively normal	CDR = 0, MMSE 24-30, WLM normal
SMC	Subjective memory concern	CDR = 0, objective tests normal, self-reported memory decline
EMCI	Early mild cognitive impairment	CDR = 0.5, mild WLM impairment
LMCI	Late mild cognitive impairment	CDR = 0.5, significant WLM impairment
AD	Alzheimer's disease (mild)	CDR = 0.5-1.0, MMSE 20-26

File	Task
train_onehot.py	Five-class classification with class-weighted CrossEntropyLoss
train_float.py	Clinical score regression with MSELoss
shap_vis_onehot.py	SHAP brain-region visualization for the classification model, including three-view slices and average heatmap NIfTI output
shap_vis_float.py	SHAP brain-region visualization for the regression model

• SHAP method: GradientExplainer, producing per-subject PNG slices and group-average .nii.gz maps

---

SimCLR_Project: Cross-Dataset Self-Supervised Pretraining

This project uses structural MRI from all five datasets as input and applies the SimCLR framework for self-supervised contrastive pretraining, learning general-purpose neuroimaging representations before extracting features for downstream tasks.

Training pipeline:

Raw MRI -> random augmentation (two views) -> 3D VGG16 encoder -> projector (4096 -> 512 -> 128) -> NT-Xent loss

3D data augmentation strategy:

• Random flipping along the X / Y / Z axes
• Random Gaussian noise
• Random intensity scaling and shifting
• Random small-angle 3D rotation (+-10°)

NT-Xent loss (normalized temperature-scaled cross entropy):

$$\mathcal{L} = -\log \frac{\exp(\text{sim}(z_i, z_j)/\tau)}{\sum_{k \neq i} \exp(\text{sim}(z_i, z_k)/\tau)}, \quad \tau = 0.5$$

File	Function
train_simclr.py	Self-supervised pretraining, saving simclr_encoder.pt
feature_extract.py	Load the encoder, extract 4096-dim features for all five datasets, and merge them into features_all.csv

---

Shared Configuration

Parameter	Value
Input size	96 x 96 x 96 (resized with trilinear interpolation)
Max epochs	1,000,000
Early stopping patience	20
GPU	0-7 (8 x H100, DataParallel)
Mixed precision	AMP (autocast + GradScaler)
Optimizer	AdamW (weight decay=1e-4)

---

Dependencies

torch >= 2.0
nibabel
pandas
numpy
scipy
scikit-learn
shap
matplotlib