National Technical University of Athens (NTUA)
Η τεχνητή νοημοσύνη στη παρακολούθηση του καρκίνου του προστάτη: από την επεξεργασία μαγνητικής τομογραφίας έως τον χαρακτηρισμό της νόσου μέσω τεχνικών μηχανικής και βαθιάς μάθησης
Abstract
dc:descriptionThis PhD Thesis focuses on the development of Artificial Intelligence (AI) methodologies across the prostate cancer (PCa) continuum, including (a) the improvement of Magnetic Resonance Imaging (MRI) through image preprocessing techniques, (b) the automated segmentation of the prostate’s peripheral zone via Deep Learning (DL) models, (c) the detection of PCa with multimodal DL models, (d) the automated segmentation of prostatic lesions employing DL models, and (e) the characterization of these lesions as clinically significant utilizing quantitative radiomics features extracted from MRI scans.More specifically, the second chapter introduces the RACLAHE methodology, which was designed to improve MRI visualization and enhance the performance of automated segmentation models for the prostate gland and its prostatic zones (Peripheral Zone - PZ, Transitional Zone - TZ). This methodology extends the existing literature by modernizing the application of image filters, such as histogram equalization, and introducing a DL model for the intelligent application onto appropriate image regions to selectively refine the region of interest. Concurrently, a methodology for the explainability of the results is developed, through which it is demonstrated that RACLAHE not only improves the performance of automated segmentation models for the prostate and its zones but also improves their interpretability. Specifically, it assists DL models in focusing on the most salient image features.The third chapter focuses on the automated segmentation of the PZ, the anatomical region where 70% of prostate cancers occur. For this purpose, the ResQu-Net model is proposed, an architecture that integrates proposed spatial attention mechanisms and squeeze-and-excitation blocks. Its comparison with six other state-of-the-art models across three different datasets highlighted its better performance in the precise delineation of the complex morphology of the PZ, while also proving, through qualitative analysis, to be more explainable than existing models, by drawing its outcomes based on the most relevant regions that constitutes the PZ. Subsequently, the fourth chapter addresses the detection of prostate cancer from MRI images, radiomic, and clinical features. Within this chapter, a multimodal Deep Learning (DL) model was developed that synergistically combines imaging, radiomic, and clinical features through a hierarchical cross-attention mechanism. To select the optimal imaging model for integration, a methodology for analyzing the expressiveness of DL features was introduced, with the proposed Weighted Expressiveness Score (WES) guiding the selection of the most robust architecture. The evaluation on a large-scale, multi-center dataset (ProstateNET) with retrospective and prospective data demonstrated that the proposed trimodal model systematically outperforms unimodal, bimodal approaches and other models present in the literature, improving the diagnostic accuracy and generalizability for prostate cancer detection. Furthermore, the fifth chapter addresses the segmentation of the cancerous lesions themselves by developing the 3D, multi-channel ProLesA-Net model. Its architecture introduces two novel multi-scale attention layers, the MSSE (Multi-Scale Squeeze and Excitation) and MSAG (Multi-Scale Attention Gate), which enable the model to detect and delineate prostatic tumors of varying sizes with high precision, placing particular emphasis on the small and medium-sized lesions that pose a significant clinical challenge. Its efficacy was confirmed through extensive evaluation against six other 3D state-of-the-art models, while an ablation study performed to prove the improvement of those attention layers (MSSE & MSAG) not only in the ProLesA-Net backbone but also in a vanilla model (U-Net). The sixth chapter completes the diagnostic pathway with the non-invasive characterization of tumors as clinically significant, presenting Simplatab. This is a comprehensive automated machine learning (AutoML) framework that leverages quantitative radiomics features from MRI scans. Simplatab is not limited to developing models with high diagnostic accuracy but also integrates advanced functionalities for data bias analysis, model vulnerability assessment, and explainable AI (XAI) through Shapley analysis, thereby improving the reliability and transparency required for the clinical adoption of such systems. Future work involves the creation of AI agents based on Large Language Models (LLMs). These agents will utilize the methodologies proposed in the dissertation as tools to establish a holistic clinical assistant in the form of a conversational agent (ChatBot). Clinicians will be able to actively interact with this agent, providing patient data and receiving responses generated by the techniques developed. In summary, a clinician could receive an enhanced image (RACLAHE), the automated segmentation of the PZ (ResQu-Net), whether a patient suffers from PCa (Multimodal Hierarchical cross-attention model), the delineation of prostatic lesions (ProLesA-Net), and a diagnosis regarding their clinical significance in a single, streamlined interaction (Simplatab).
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Zaridis, Dimitrios
- Ζαρίδης, Δημήτριος
Subjects
dc:subject × 32- Καρκίνος του προστάτη
- Ουρολογία
- Ακτινολογία
- Τμηματοποίηση ιατρικής εικόνας
- Ανίχνευση
- Βαθιά μάθηση
- Μηχανική μάθηση
- Τεχνητή νοημοσύνη
- Συνελικτικά νευρωνικά δίκτυα
- Πολυτροπική μάθηση
- Επεξεργασία εικόνας
- Βιοϊατρική τεχνολογία
- Ανάλυση εικόνας
- Prostate cancer
- Urology
- Radiology
- Medical image segmentation
- Detection
- Deep learning
- Machine learning
- Artificial intelligence
- Convolutional neural networks
- Multimodal learning
- Image processing
- Biomedical technology
- Image analysis
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Βιοπληροφορική
- Computer and Information Sciences
- Natural Sciences
- Bioinformatics
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60809