Title Evaluating and enhancing semi-supervised learning algorithms for pancreatic cancer segmentation in ct images /
Translation of Title Kasos vėžio segmentavimo KT vaizduose įvertinimas ir tobulinimas pusiau prižiūrimų mokymosi algoritmų.
Authors Ahmed, Md Istiaque
Full Text Download
Pages 52
Keywords [eng] Semi-Supervised Learning, Pancreatic Cancer, CT Image Segmentation, MeanTeacher, MixMatch, Deep Learning, Medical Imaging, Limited Labeled Data.
Abstract [eng] Semi-supervised learning (SSL) offers a promising avenue to address the scarcity of labeled data in medical image segmentation, a critical challenge for tasks like detecting pancreatic cancer from Computed Tomography (CT) scans. This Master’s thesis evaluates the efficacy of two prominent SSL algorithms, Mean Teacher and MixMatch, for 2D pancreatic CT image segmentation. The study compares their performance against a robust supervised U-Net baseline when trained with a severely limited set of labeled data and a larger pool of unlabeled data derived from the Medical Segmentation Decathlon Pancreas-CT dataset. Experiments revealed that while the supervised baseline achieved strong performance with limited labeled data after extensive training, neither Mean Teacher nor MixMatch, under the tested configurations, demonstrated a significant improvement over this baseline. Key challenges were identified, including teacher model instability in the Mean Teacher framework and the generation of poor-quality pseudo-labels leading to performance degradation in the MixMatch approach. These issues were particularly pronounced in the low-data regime. This research contributes a rigorous evaluation of these SSL techniques in a challenging clinical application, highlighting practical limitations concerning model stability and pseudo-label reliabil- ity with scarce annotations. The findings underscore the complexities of effectively leveraging unla- beled data for this task and provide insights for future development of more robust semi-supervised methods for medical imaging.
Dissertation Institution Vilniaus universitetas.
Type Master thesis
Language English
Publication date 2025