Adversarial learning for MRI reconstruction and classification of cognitively impaired individuals

This work is published in IEEE Access (https://doi.org/10.1109/ACCESS.2024.3408840).

Introduction

This is the repo for the model proposed in the paper. Briefly, the model performs medical image (scans) reconstruction. The result shows that the reconstructed scans not only have better image quality, but also improves the prediction accuracy of MRI progression (sMRI vs pMRI).

The model is trained on a dataset from ADNI and evaluated on an exteranl dataset from NACC.

Image quality metrics:

(a) Results on ADNI cohort test partition

Input Images	CNR	SNR	SSIM	BRISQUE	PIQE
Original	2.111±0.274	0.686±0.052	-	43.513±1.594	41.936±2.062
Diced	0.726±0.131	0.537±0.046	0.319±0.065	43.567±0.333	78.307±5.322
GAN-VAN	2.235±0.299	1.857±0.214	0.544±0.051	42.485±0.741	72.028±3.068
GAN-NOV	1.934±0.335	1.437±0.395	0.580±0.059	42.246±0.563	68.120±1.351

(b) Results on NACC cohort

Input Images	CNR	SNR	SSIM	BRISQUE	PIQE
Original	2.065±0.294	0.676±0.074	-	44.292±2.940	43.182±6.298
Diced	0.696±0.156	0.526±0.074	0.348±0.108	43.742±1.256	78.442±4.695
GAN-VAN	2.193±0.359	1.821±0.230	0.523±0.105	42.563±1.128	72.133±4.024
GAN-NOV	1.790±0.363	1.432±0.534	0.553±0.116	42.505±1.066	68.176±2.079

MRI progression prediction performance:

(a) Results on ADNI cohort test partition

Input Images	Accuracy	Precision	F1-score	MCC
Original	0.794±0.033	0.795±0.032	0.785±0.036	0.516±0.081
Diced	0.703±0.064	0.681±0.082	0.665±0.080	0.254±0.145
GAN-VAN	0.709±0.069	0.711±0.084	0.653±0.079	0.270±0.130
GAN-NOV	0.709±0.042	0.697±0.062	0.694±0.060	0.303±0.133

(b) Results on NACC cohort

Input Images	Accuracy	Precision	F1-score	MCC
Original	0.675±0.006	0.683±0.004	0.671±0.007	0.358±0.010
Diced	0.597±0.021	0.624±0.021	0.573±0.028	0.219±0.042
GAN-VAN	0.589±0.025	0.625±0.021	0.556±0.039	0.211±0.047
GAN-NOV	0.640±0.017	0.650±0.014	0.634±0.019	0.289±0.030

See the paper for additional information

Quick start

1. CUBLAS_WORKSPACE_CONFIG=:4096:8 python rcgan_main.py (Train and generate scans for single G)
2. CUBLAS_WORKSPACE_CONFIG=:4096:8 python rcgans_main.py (Train and generate scans for multiple G)
3. CUBLAS_WORKSPACE_CONFIG=:4096:8 python classifier_main.py (Evaluate prediction performance using CNN)
4. (optional, in plot/) python plot.py
5. (optional, image quality) python image_quality.py
6. (optional, MCC) python matrix_stat.py

Environments

1. Install python3
2. Install the environments.yml (Anaconda environment)
3. (optional, image quality) Install matlab for python (if standard method not work, try: sudo python setup.py install --prefix="/home/xzhou/anaconda3/envs/py36/")

Data Preprocessing

The data preprocessing follows a similar procedure as in this work: (https://github.com/vkola-lab/mri-surv-dev)

Hyper-parameter Tuning

The json files that contains 'config' name (i.e. config.json) are the files that can be used to modify most of the hyperparameters.