Multi-class classification of Alzheimer's disease detection from 3D MRI image using ML techniques and its performance analysis

Alzheimer's disease is a prevalent kind of syndrome; critical to diagnose in its early stages causes the patient forgets everything in its later stages. In this work, we proposed a design for early diagnosis of Alzheimer's disease; where a multi-class classification system has been implemented which detects AD and classifies the level of disease as Normal, Mild and Severe. The proposed approach starts with mapping the brain's anatomical parts hippocampal, white matter and grey matter and respective volumes are calculated from 3D MRI images. The image segmentation and calculation of volume are done with two software; Analyze Direct and ITK Snap. Calculated volumes of the anatomical parts along with other features like age, gender and MMSE score are fed to different machine learning algorithms for Alzheimer's detection as well as its severity. The extracted features are also fused randomly in all possible ways for further analysis using ML classifiers. The ML algorithms used are random forest, gradient boost, decision tree and KNN. The proposed approach is tested with two sets of data; OASIS dataset and ADNI dataset. Classifier's performance is analyzed based on sensitivity, F1 Score, accuracy and precision for ML classifiers. Random forest is giving the highest accuracy of 99% for white matter volume using OASIS dataset and when all three volumes of hippocampal, white matter and grey matter are fused giving 98% accuracy. For ADNI data set using white matter volume, we are getting 92% accuracy for gradient boost classifier and after fusing all three volumes also getting 92% accuracy. Gradient boost gives an accuracy of around 91% for both databases.