Development of an ensemble deep learning model for brain stroke detection and classification in MRI scans

Lee, Ding Kuan (2026) Development of an ensemble deep learning model for brain stroke detection and classification in MRI scans. Final Year Project, UTAR.

PDF Download (5Mb)

Abstract

Stroke is the leading cause of death and disability globally. More than half of all new strokes are ischemic strokes, which result from blood clotting and obstructed blood supply to the brain. Timely and accurate detection is essential, but interpreting magnetic resonance imaging (MRI) scans is time-consuming and often inaccurate. This project created NeuroScan, an ensemble deep learning web application for predicting ischemic stroke from brain MRI scans. Six pre-trained deep learning models were fine-tuned: ResNet50, ResNet101, DenseNet121, DenseNet169, EfficientNet-B3, and Vision Transformer (ViT). The models were fine-tuned using a dataset of 15,120 MRI scans with labels from Hospital Pengajar Universiti Putra Malaysia (HPUPM), with 38 and 34 scans used for validation and testing, respectively. The ensemble learning methods tested were simple averaging, weighted averaging, hard voting, and stacking. Three ensemble methods and the ResNet50 model achieved perfect accuracy on the test set (1.000). Grad-CAM and Attention Rollout were adopted to provide visual explanations of the predictions. The models were deployed in a Flask web application with Firebase Authentication and Firestore, incorporating three-layer role-based access control for patients, doctors, and administrators. The platform enabled the complete workflow for uploading an MRI scan, review by a doctor, assistance from an AI chatbot (Claude Haiku), two types of notifications (in-app and email), and PDF report export. Security features included Content Security Policy (CSP) headers, rate limiting, session timeouts, and input validation. Docker was used to deploy the application on Hugging Face Spaces. No critical issues were found in 38 functional tests, 90 cross-browser tests, and 10 security tests. All four project goals were fully achieved.

Actions (login required)