Πλοήγηση ανά Συγγραφέας "Marias, Konstantinos"

Τώρα δείχνει 1 - 1 of 1
  • Μικρογραφία εικόνας
    Τεκμήριο
    Development and optimization of AI models based on medical images
    (ΕΛΜΕΠΑ, Σχολή Μηχανικών (ΣΜΗΧ), ΠΜΣ Μηχανικών Πληροφορικής, 2026-03-03) Oikonomou, Nikolaos; Οικονόμου, Νικόλαος; Marias, Konstantinos; Μαριάς, Κωνσταντίνος
    Prostate cancer (PCa) and therapy-induced cardiac toxicity represent two critical challenges in precision oncology: the former as one of the most frequent male cancers worldwide and the latter as severe side effect of breast cancer treatment that significantly impacts patient outcomes. Artificial Intelligence (AI) offers promising solutions to both problems, however its adoption is limited by heterogeneous imaging data and the lack of dedicated echocardiographic-based models for cardiotoxicity prediction. This master’s thesis addresses these gaps through two complementary studies using datasets from the EU-funded ProCAncer-i and CARDIOCARE projects. For PCa management, bi-parametric MRI scans (T2W, ADC, DWI) were employed to train four 3D deep learning models under different harmonization strategies, including z-score, mean and histogram normalization and the combination of z-score and N4 bias field correction. For the cardiotoxicity risk prediction, 3D echocardiographic videos acquired from breast cancer patients were employed, using 3D ResNet18 and I3D architectures to predict cardiotoxicity occurrence within one year of the baseline examination. The results show the significance of harmonization techniques in enhancing AI models generalization ability on heterogeneous MRI data. Specifically, in PCa management, ResNet18 trained on histogram-normalized 24x192x192 MR images, achieved an AUC of 71.28%, compared to 63.52% with raw data. In cardiotoxicity risk prediction, ResNet18 outperformed I3D and EchoNet pre-trained models, achieving a mean AUC of 64.52% using the A4C view at low resolution data. Taken together, these findings highlight the value of harmonization for PCa AI models and demonstrate the feasibility of video-based deep learning models as a supportive tool for cardiotoxicity monitoring.