AN AI BASED ALZHEIMER’S CLASSIFICATION SYSTEM USING MRI IMAGES AND XAI

AZIE, CHIMDIUME (2026) AN AI BASED ALZHEIMER’S CLASSIFICATION SYSTEM USING MRI IMAGES AND XAI. Other thesis, GODFREY OKOYE UNIVERSITY, ENUGU,.

Text AZIE CHIMDIUME PROJECT.pdf Download (3MB)

Abstract

Alzheimer's disease is the major causative factor of dementia in the world with more than 55 million people suffering from it, and the numbers have been increasing drastically in developing nations such as Nigeria, where access to diagnostic services is severely lacking. The early diagnosis of Alzheimer's disease is paramount in ensuring the treatment of the condition through clinical interventions; however, the current method of diagnosing Alzheimer's is very subjective as it uses brain MRI scans interpreted by radiologists, a resource which is in critically short supply in hospitals within low-income settings. The current research aims to provide an end-to-end artificial intelligence based Alzheimer's disease detection and hospital management solution which combines deep learning classifiers and explainable artificial intelligence techniques. The proposed model uses the EfficientNet-B0 Convolutional Neural Network architecture, and its training was achieved through transfer learning on the OASIS Alzheimer's MRI data set consisting of over 90,000 brain MRI images of four different disease stages, namely, NonDemented, VeryMildDemented, MildDemented, and ModerateDemented. Training was done using class weighting to solve the issue of extreme dataset imbalance, while Gradient-weighted Class Activation Mapping (GradCAM) was employed as an after-the-fact explainability approach to create heat maps showing the regions of the brain which affected predictions the most. The validation accuracy obtained during training reached 98.66% in the 19th epoch out of a total of 25 training epochs, with the weighted average F1-score of the model reaching 0.8126 on the unseen test data set, consistently emphasizing clinically important parts of the brain such as the hippocampus and ventricles through its Grad-CAM. The trained model was used in a full-stack hospital management web application developed using Flask and vanilla JavaScript, with functionality such as role-based access control for Admin and Doctor roles, patient registration, uploading of MRIs, storing the results, tracking the history of scans and even transferring of patients, all of which were then hosted on a cloud hosting platform. This makes the model clinically deployable since it can be used in hospitals even with regular computational hardware, an important contribution towards early diagnoses of AD in limited resources areas.

Item Type:	Thesis (Other)
Subjects:	Q Science > Q Science (General)
Divisions:	Faculty of Computing And Information Technology (FACIT)
Depositing User:	Nnenna Ayo
Date Deposited:	24 Jul 2026 10:26
Last Modified:	24 Jul 2026 10:26
URI:	http://eprints.gouni.edu.ng/id/eprint/5911

Actions (login required)

View Item