# Detection of Alzheimer's Disease Using Graph-Regularized Convolutional Neural Network Based on Structural Similarity Learning of Brain Magnetic Resonance Images

**Authors:** Kuo Yang, Emad A. Mohammed, Behrouz H. Far  
**arXiv:** [2102.13517](https://arxiv.org/abs/2102.13517)  
**Date:** 2021-02-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2102.13517v1/

## Key points

- Authors present a novel method for detecting Alzheimer's disease using a Graph-Regularized Convolutional Neural Network based on Structural Similarity Learning of Brain MRIs
- Study aims to leverage structural similarity between MRIs by representing it as a graph
- Construct a similarity graph utilizing embedded features from input images categorized into ND, VMD, MD, and MDTD classes
- Experiment with dimension-reduction and clustering algorithms to create an optimal similarity graph using cosine distance as a measure
- Employ the constructed similarity graph to sample training data for CNN model training
- Similarity graph serves as regularizer in the loss function of CNN model
- Extensive experiments conducted using pre-trained CNN models show superior performance on testing dataset with accuracy of 0.986, AUC of 0.998, and F1 measure of 0.987
- Emphasize significance of classification results in enhancing AD detection accuracy and release all code used in experiments for reproducible research

## Summary

Authors Kuo Yang, Emad A. Mohammed, and Behrouz H. Far present a novel method for detecting Alzheimer's disease (AD) using a Graph-Regularized Convolutional Neural Network based on Structural Similarity Learning of Brain Magnetic Resonance Images (MRIs). Their study aims to leverage the structural similarity between MRIs by representing this similarity as a graph. The authors construct a similarity graph utilizing embedded features from input images categorized into Non-Demented (ND), Very Mild Demented (VMD), Mild Demented (MD), and Moderated Demented (MDTD) classes. They experiment with various dimension-reduction and clustering algorithms to create an optimal similarity graph that captures similarities within the same class images using cosine distance as a measure. To train their model, they employ the constructed similarity graph to sample training data for a convolutional neural network (CNN). The similarity graph serves as a regularizer in the loss function of the CNN model, aiming to minimize the distance between input images and their k-nearest neighbors in the graph while also minimizing categorical cross-entropy loss between predicted image labels and actual class labels. Extensive experiments are conducted using several pre-trained CNN models, with results compared against recent methods. The proposed method demonstrates superior performance on the testing dataset, achieving an accuracy of 0.986, area under receiver operating characteristics curve of 0.998, and F1 measure of 0.987. This signifies an improvement in prediction accuracy compared to existing methods. The authors emphasize the significance of their classification results in enhancing AD detection accuracy and release all code used in their experiments to promote reproducible research in this field. Their innovative approach showcases the potential for leveraging structural similarity learning in MRI analysis for improved Alzheimer's disease detection outcomes.