Doktorandské kolokvium KAI - Jafari Fatana (24.2.2025)
v pondelok 24.2.2025 o 13:10 hod. v miestnosti I/9
Prednášajúci: Jafari Fatana
Názov: Automatic Classification and Diagnosis of Multiple Sclerosis Lesions(White Matter) From Brain MR Images Based on Deep Learningand Explainable AI Methods
Termín: 24.2.2025, 13:10 hod., I/9
Abstrakt:
This research investigates the detection of white matter lesions in multiple sclerosis (MS) using brain magnetic resonance imaging (MRI). MS is a chronic autoimmunedisease affecting the central nervous system, leading to symptoms such as muscleweakness, dizziness, fatigue, and vision loss. Early diagnosis is crucial for timelytreatment and disease progression prevention. While AI-based systems have shownsuccess in identifying MS lesions, their integration into clinical practice is limitedbya lack of transparency and explainability, undermining clinician trust. Explainableartificial intelligence (XAI) addresses this issue by providing interpretablejustifications for model predictions, resolving the "black-box" probleminherent inmany AI systems.
The primary goal of this research is to develop novel, fully automated methods for detecting white matter lesions in MS by integrating post hoc explanation techniques into deep learning models, including local interpretable model-agnostic explanations (LIME) and Gradient-weighted Class Activation Mapping (Grad-CAM), to enhancethe interpretability of AI predictions. Deep Convolutional Neural Networks (CNNs) were trained on a dataset of 514 brain MRI scans (345 positive and 169 negative) inDICOM format, obtained from the radiology department of Cyril and Methodius Hospital in Bratislava.
Prior to training, several pre-processing techniques were applied to improve imagequality. Firstly, we converts a series of DICOM images into a single NIfTI file whichsimplifying processing and compatibility with neuroimaging tools. this canbeachieved using the SimpleITK library to read the DICOM series and save it as a 3DNIfTI file. Secondly, we applied Skull Stripping technique using (FSL's BET) whichis FMRIB Software Library with the Brain Extraction Tool for removing non-braintissues, such as the skull and scalp, from brain imaging data to focus analyses on brainstructures. Additionally, implemented Bias Field Correction (N4ITK) methodtocorrects MRI intensity non-uniformities caused by field inhomogeneities, enhancingimage quality; and so on.