Doktorandské kolokvium KAI - Fatana Jafari (16.11.2026)
v pondelok 16.11.2026 o 13:10 hod. v miestnosti I 9
Prednášajúci: Fatana Jafari
Názov: Explainable Deep Learning for White Matter Lesion Analysis in Brain MRI: From 3D Architecture Evaluation to Hybrid CNN–Transformer Models
Termín: 16.11.2026, 13:10 hod., I 9
Abstrakt:
Detecting white matter lesions (WMLs) in brain MRI can be challenging, especially when lesions are small and have low contrast. This research explores deep learning methods for WML detection and analysis using 3D FLAIR MRI scans. Several 3D deep learning architectures were trained and evaluated using a dataset collected at St. Cyril and Methodius Hospital in Bratislava. Explainable AI (XAI) methods were also used to better understand model predictions and identify the image regions that influence them.
Building on these experiments, the research focuses on developing a hybrid model that combines Xception and Vision Transformer to capture both local image features and broader contextual information. Different patch sizes are being explored to improve the detection of small lesions. Grad-CAM, SHAP, and LIME are also being investigated to understand model decisions and compare the highlighted regions with lesion annotations. The aim is to develop deep learning methods that make white matter lesion detection more accurate and easier to interpret.