| Abstract [eng] |
This master’s thesis investigates semi-supervised learning methods for the identification of pa ncreatic cancer in computed tomography images. Pancreatic cancer remains one of the most chal lenging oncological diseases to diagnose, making the application of artificial intelligence for early detection highly valuable. However, the limited availability of annotated medical images const rains the development of accurate learning models. To address this, the study explores the use of semi-supervised learning techniques that can leverage both labeled and unlabeled data. The rese arch begins with a literature review, analyzing the fundamentals of semi-supervised learning, its historical development, application challenges in medical image analysis, and the key algorithms considered, the Mean Teacher method, its uncertainty-based extension, and the FixMatch appro ach. Publicly available computed tomography image datasets were also evaluated against several criteria, including clinical relevance, annotation quality, accessibility, image diversity, and comp lexity, to select the most suitable datasets for experimentation. In the experimental phase, the three selected methods were applied to pancreas segmentation tasks, aiming to assess their effectiveness under conditions of limited labeled data. All models were trained using the same network architec ture and evaluated with consistent metrics, the Dice coefficient and the 95th percentile Hausdorff Distance. The results demonstrated that all semi-supervised learning methods significantly impro ved segmentation quality. Overall, the findings confirm that semi-supervised learning presents a promising solution for enhancing medical image analysis, particularly in scenarios where annotated data are scarce. |
