National Technical University of Athens (NTUA)
Αυτοματοποιημένη διάγνωση και πρόγνωση παθολογιών μέσω απεικονιστικών εξετάσεων: προσεγγίσεις μηχανικής μάθησης στο οστεοσάρκωμα, στην αλλεργική δερματίτιδα εξ επαφής, και στο αδενοκαρκίνωμα του παγκρέατος
Abstract
dc:descriptionIn an era characterized by groundbreaking advancements in technology and healthcare, the convergence of machine learning and medicine has emerged as a contemporary frontier poised to reshape the landscape of healthcare. Central to this convergence is the domain of medical image analysis, where the synergy between computational prowess and clinical expertise holds immense promise. Amid escalating demands for precise and effective diagnostic and prognostic tools, this doctoral thesis endeavors to lead the way with novel approaches harnessing the potential of machine learning to diagnose and prognose pathologies across three distinct imaging modalities. Specifically, the research focuses on the analysis of histological images featuring osteosarcoma, multispectral dermoscopy images indicative of allergic contact dermatitis, and preoperative CT scans of patients with pancreatic ductal adenocarcinoma. In the first study, the research delves into the diagnosis of histological osteosarcoma images. A comparative methodological approach is adopted to evaluate modern deep neural networks, trained on a publicly available dataset, for the automated detection of viable and necrotic tumors, as well as healthy tissue. The research emphasizes the importance of careful selection of different network architectures, their depth, and the dimensions of the input images, achieving optimal overall accuracy of 91% using the MobileNetV2 network and an input image size of 256×256 pixels. The second study focuses on the diagnosis of Allergic Contact Dermatitis (ACD) using deep learning models, different pre-processing schemes and image modalities. A deep learning approach is employed, incorporating a pre-processing technique to highlight regions of interest while preserving the overall information of the images. Additionally, the study explores the utility of different image modalities and the combination thereof to achieve optimal diagnostic accuracy. The results showed that the combination of the color and redness image modalities produce optimal results, achieving an accuracy greater than 91% in detecting the allergens to which the patient manifested sensitivity to. This research represents one of the earliest efforts in automated diagnosis of ACD, paving the way for the development of advanced tools and techniques that expedite diagnosis while reducing clinicians’ workload. In the third study, a fully automated approach is proposed for the prognostication of pancreatic adenocarcinoma, utilizing preoperative CT scans of patients undergoing pancreatectomy. This approach entails the use of deep learning techniques for tumor and pancreas segmentation, as well as the extraction of radiomics and clinical features to calculate overall survival using traditional machine learning techniques. The proposed approach achieved a Harrell’s Concordance Index (C-Index) of 0.731, outperforming existing methods based on clinical parameters, such as TNM staging, while highlighting the importance of specific radiomics features, especially the LAHGLE.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Vezakis, Ioannis
- Βεζάκης, Ιωάννης
Subjects
dc:subject × 27- Μηχανική μάθηση
- Ιατρικές εικόνες
- Οστεοσάρκωμα
- Αλλεργική δερματίτιδα εξ επαφής
- Αδενοκαρκίνωμα παγκρέατος
- Διάγνωση
- Πρόγνωση
- Ιστοπαθολογία
- Δερματοσκόπηση
- Επιδερμιδικές δοκιμασίες επιθέματος
- Αξονική τομογραφία
- Machine learning
- Medical imaging
- Osteosarcoma
- Allergic contact dermatitis
- Pancreatic adenocarcinoma
- Diagnosis
- Prognosis
- Histology
- Dermoscopy
- Patch testing
- Computed tomography
- Βιοϊατρική Μηχανική
- Επιστήμες Μηχανικού και Τεχνολογία
- Medical Engineering
- Engineering and Technology
- Biomedical Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/55222