Abstract
Alzheimer’s disease affects multiple areas of the brain, one of which is the Corpus Callosum—a compact, yet critical structure whose size and shape make its modelling quite challenging. Specifically, I was interested in investigating whether Alzheimer’s can be identified using only this area. In my initial experiments, I utilized the OASIS-2 dataset [1], which holds 350 brain images labelled as Alzheimer’s, MCI, or normal. Several CNN models, i.e., VGG, ResNet, and EfficientNet, were trained using various preprocessing strategies such as grayscale conversion, data normalization, contrast stretching, data augmentation, and SMOTE. Yet, low model performance was caused by insufficient data and the compact size of the Corpus Callosum. To address this, I moved to the ADNI dataset [2], comprising more than 3,000 brain scans. I had to transform the 3D. niftii scans into 2D slices and tried various resolutions and CNN architectures. I also tried applying a 3D CNN to unsegmented scans, although this gave mixed results due to added brain tissue causing noise and decreasing precision. Isolating the Corpus Callosum became apparent as essential for enhancing performance.
To do so, I employed a U-Net segmentation model [3] for extracting Corpus Callosum from the scans. Classification was limited to this segmented area, and this improved precision to about 80%. This result proves the viability of using the Corpus Callosum as an accurate biomarker for detecting Alzheimer’s disease in support of using specific brain areas for disease diagnosis.