National and Kapodistrian University of Athens
Use of machine learning algorithms to optimize diagnostic methods using magnetic resonance imaging of adult patients
Abstract
dc:descriptionThis PhD research advances neuroimaging artificial intelligence from conceptual prototypes toward clinically deployable systems designed for real-world radiology workflows. It introduces a holistic framework that balances high diagnostic accuracy with interpretability, modular architecture with cohesive functionality, and robustness with adaptability. By addressing practical constraints—including multi-pathology coverage, multimodal MRI inputs, raw data processing, clinician-oriented outputs, and readiness for embedded deployment—the proposed system functions as a trustworthy copilot for radiologists, aiming to reduce diagnostic delays, alleviate cognitive load, and improve consistency in clinical decision-making.Development and evaluation are based on a carefully curated multiparametric brain MRI dataset comprising 150 examinations from a single 3T academic center. The dataset captures realistic clinical heterogeneity across seven major pathologies and includes six complementary MRI sequences (FLAIR, T2, T1ce, DWI, ADC, and T1) in both 2D and 3D formats. Rigorous quality control and reproducible preprocessing pipelines enable realistic benchmarking while acknowledging limitations such as class imbalance and single-center bias, thereby establishing a foundation for future multi-center validation. The proposed cascaded deep learning pipeline combines transfer learning–based slice-level classification with U-Net–based segmentation, achieving high diagnostic accuracy and spatially explicit lesion delineation. This dual strategy optimizes specificity for normal examinations and sensitivity for abnormality detection, while enhancing interpretability through volumetric summaries and lesion contextualization. Systematic investigation of modality selection and architectural design further demonstrates the advantages of multi-sequence fusion, 3D contextual information, and domain adaptation for clinical robustness. Additional contributions include the application of federated learning in neuroimaging, demonstrating improved performance, privacy preservation, and scalability for collaborative multi-center deployment. Furthermore, the integration of multimodal MRI radiomics with advanced classifiers such as XGBoost provides a scalable and interpretable framework for broad pathology differentiation. Overall, this thesis presents a clinically relevant, interpretable, and scalable neuroimaging AI system that supports triage, prioritization, and improved patient care, contributing to narrowing the gap between emerging AI methods and their routine clinical adoption.
Degree
thesis:*- Grantor dc:publisher
- National and Kapodistrian University of Athens
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Stathopoulos, Ioannis
- Σταθόπουλος, Ιωάννης
Subjects
dc:subject × 20- Απεικόνιση μαγνητικού συντονισμού
- Παθολογίες Εγκεφάλου
- Τεχνητή νοημοσύνη
- Βαθειά Μηχανική Εκμάθηση
- Επιστήμη δεδομένων
- Magnetic resonance imaging
- Brain pathophysiology
- Artificial intelligence
- Deep learning
- Data science
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Επιστήμες Υγείας
- Ιατρική και Επιστήμες Υγείας
- Ιατρική Πληροφορική
- Computer and Information Sciences
- Natural Sciences
- Health Sciences
- Medical and Health Sciences
- Medical Informatics
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/10442/hedi/61098