Image-based regression of disease biomarkers using deep learning modelsDULAR, LARA (Avtor) Špiclin, Žiga (Mentor) Brain ageBrain age gapT1-weighted MRIImaging biomarkersDeep learningDeep regressionEvaluation protocolMedical imaging–derived biomarkers have emerged as a promising, noninvasive approach for characterizing disease-related structural and functional changes and for supporting diagnosis, prognosis, and patient monitoring. Brain age estimation from structural magnetic resonance imaging (MRI) is one such imaging biomarker for neurological health. The difference between predicted brain age and chronological age, known as the brain age gap (BAG), has been associated with several neurological and systemic diseases. However, the practical use of deep learning–based brain age prediction remains limited by inconsistent preprocessing pipelines, lack of standardized evaluation protocols, and reduced performance when models are applied to data from new scanners or populations. This thesis investigates image-based regression of disease biomarkers using deep learning models, with a primary focus on brain age estimation from T1-weighted MRI. The work addresses key methodological challenges through four main contributions. Frist, the thesis introduces the Brain Age Standardized Evaluation (BASE) framework, a unified protocol for benchmarking brain age models across multi-site, unseen-site, test–retest, and longitudinal datasets using a comprehensive set of performance metrics and statistical analyses. Using this framework, models trained on full-resolution 3D images were shown to outperform 2D slice-based approaches in both accuracy and consistency. Second, the impact of MRI preprocessing on prediction accuracy and generalization was systematically analyzed across multiple deep learning architectures. The results demonstrate that extensive preprocessing significantly improves prediction accuracy and cross-site generalization. Third, the thesis investigates strategies for improving model performance on small or domain-shifted datasets, demonstrating that techniques such as data augmentation, transfer learning, and bias correction can mitigate performance degradation caused by scanner differences and limited training data. Finally, the clinical relevance of brain age estimation is evaluated by analyzing BAG across cohorts with neurological and systemic diseases, including multiple sclerosis, Parkinson’s disease, mild cognitive impairment, Alzheimer’s dementia, and type II diabetes. The results show disease-specific increases in BAG and highlight differences in biomarker behavior across model architectures.20262026-03-17 13:40:15Doktorsko delo/naloga180824sl