Ελληνικό Μεσογειακό Πανεπιστήμιο
Prostate cancer assessment based on multimodal data and deep learning methods
Abstract
dc:descriptionProstate cancer is the second most frequently diagnosed malignancy in men worldwide, making timely diagnosis essential for treatment decisions and patient outcomes. Diagnosis typically combines prostate-specific antigen (PSA) testing, magnetic resonance imaging (MRI), and, when indicated, biopsy. However, PSA lacks specificity and can cause overdiagnosis; biopsy is invasive and susceptible to sampling error; and standardized radiological assessment shows substantial inter-reader variability across institutions and expertise levels. Robust, generalizable models are further limited by scarce large, curated datasets with lesion-level annotations and by variation in MRI protocols and scanner vendors. Deep learning can learn discriminative imaging patterns directly from data. This capacity, together with these challenges, motivates multimodal, annotation-efficient, and foundation-model-based approaches that move beyond isolated lesion analysis towards clinically applicable patient-level assessment. This thesis first critically and systematically reviews generative deep learning for oncological MRI augmentation, identifying key evaluation strategies, best practices, and open challenges. It shows that image realism and conventional quality measures alone cannot establish the clinical value of synthetic data; evaluation on defined downstream tasks is essential. The thesis then develops and validates annotation-efficient architectures for patient-level prostate cancer assessment from biparametric MRI using large multicenter retrospective data and prospective cohorts. It investigates whether complementary MRI inputs, integrated through cross-attention and combined with routinely available demographic and clinical variables, improve classification of prostate cancer and clinically significant disease, while assessing performance across patient subgroups, scanners, and institutions. The proposed pipelines were benchmarked against clinical PI-RADS assessments. The work also adapts medical vision foundation models using sequence-specific, parameter-efficient techniques, enabling scalable patient-level analysis without costly lesion annotations. Collectively, the studies demonstrate the potential of multimodal, annotation-efficient artificial intelligence and provide large-scale retrospective and prospective validation, patient- and acquisition-related subgroup analyses, calibration, interpretability, and computational-efficiency evaluation, strengthening evidence for future clinical translation.
Degree
thesis:*- Grantor dc:publisher
- Ελληνικό Μεσογειακό Πανεπιστήμιο
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Dimitriadis, Avtantil
- Δημητριάδης, Αβταντίλ
Subjects
dc:subject × 18- Βαθιά μάθηση
- Μαγνητική τομογραφία
- Καρκίνος του προστάτη
- Πολυτροπική μάθηση
- Διασταυρούμενη προσοχή
- Θεμελιώδη μοντέλα
- Deep learning
- Magnetic resonance imaging (MRI)
- Prostate cancer
- Multimodal learning
- Cross-attention
- Foundation models
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62861