This is a project that aims to compare multiple classification methods including KNN and CNN network in Alzheimer’s disease diagnosis, reach higher accuracy with only single-angled MRI dataset and use GAN network to reconstruct brain MRI photos of Alzheimer’s disease patients, which are possible to then be fed back to the classification network to test its result.

Our dataset comes from Kaggle-Alzheimer MRI Preprocessed Dataset. Methodologies include Naive Bayes, Decision Tree, SVM, KNN, Logistic Regression, Random Forest, CNN and GAN. It turns out common classification methods like Random Forest and Logistic Regression can reach approximately 90% accuracy already.

CNN is proposed to further increase the accuracy. BY adjusting our models and parameters, we reach around 98% accuracy in MRI classification with the following structure.

Statistics for CNN prediction:

Furthermore, we aim to let neural networks learn the patterns of MRI. GAN is used to reproduce similar MRI images based on the training set. The basic algorithm is as follows:

The images we reproduce seem also promising, with a high likelihood to original ones.

For more details, codes can be found in the .ipynb files and Appendix.pdf. For the report, please refer to Alzheimer_MRI_Diagnosis_and_Reproduction.pdf.

Contributors: Yifan Li, Zimu Gong, Ming Wang