| Meno: | Adam |
|---|---|
| Priezvisko: | Lopaąka |
| Názov: | Generative Neural Networks for Data Augmentation in Brain MRI Classification Based on the Presence of the White Matter Lesion |
| Vedúci: | doc. RNDr. Zuzana Černeková, PhD. |
| Rok: | 2026 |
| Kµúčové slová: | brain MRI, white matter lesions, data augmentation, diffusion models, medical image classification |
| Abstrakt: | This thesis focuses on the use of generative neural networks for data augmentation in brain magnetic resonance imaging classification based on the presence of white matter lesions. White matter lesions are important imaging findings in neurological diseases, but their automatic analysis is difficult because annotated medical data are limited, lesions have variable appearance, and MRI data require careful preprocessing. The main goal of this work is to investigate whether synthetic brain MRI volumes generated by a diffusion model can improve binary classification of MRI scans into lesion-positive and lesion-negative cases. The thesis first reviews the medical background of white matter lesions, MRI-based analysis, data augmentation methods, and current generative approaches in medical imaging. Then, a preprocessing pipeline for FLAIR MRI volumes is designed, and a three-dimensional denoising diffusion probabilistic model is trained using the MONAI framework. The generated data are evaluated by visual inspection, MS-SSIM, FID, downstream classification performance, and expert radiological assessment. The results show that the proposed DDPM model is able to generate structurally plausible brain MRI volumes. Synthetic augmentation improved several classification metrics across tested deep learning models and helped make the classifiers more robust on real test data. A radiologist also assessed most generated lesion-positive volumes as realistic, especially for small scattered white matter lesions. However, some synthetic volumes still showed limitations, mainly excessive hemispheric asymmetry and simplified lesion structure. Overall, the thesis shows that diffusion-based synthetic MRI augmentation is a promising support tool for white matter lesion classification, but it should complement, not replace, real clinical data. |
Súbory diplomovej práce:
| diplomova_praca_en.pdf |
Súbory prezentácie na obhajobe: