Preprocess and Classification of Amyloid-β Pathological Images Using Vision Transformer for the Diagnosis of Alzheimer’s Disease

The diagnosis of Alzheimer’s disease can make use of visual applications of machine learning. Samples of human brain with the disease tend to have appearances of Amyloid Beta protein, while those that do not have the disease typically do not have such protein. With DAPI staining, the marks of Aβ proteins become visually distinctive. In this experiment sample images obtained using DAPI staining and fluorescence microscopy are collected as a dataset. Images in the dataset are then preprocessed with different methods and used to train computer vision classifiers based on the vision transformer to classify new images of human brain samples into two classes based on those patterns: significant presence of Amyloid-β protein or pathologically insignificant presence of Amyloid-β protein. The author finds that a classifier model based on ViT can reach a high accuracy (98% or up to 100% depending on the dataset) and can be deployed towards assisting the diagnosis of Alzheimer’s Disease. In addition, for the ViT classifier, the most effective steps of preprocessing are increasing the brightness of the images or turning the images into grayscale, depending on the specific conditions of the dataset.