Loading...

Slice-Level Weakly Supervised Learning in 3D Medical Image Segmentation

Bitarafan, Adeleh | 2024

0 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 58297 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Soleymani Baghshah, Mahdieh
  7. Abstract:
  8. Accurate segmentation of 3D medical images is a fundamental task in image processing and medical data analysis, playing a critical role in disease diagnosis, treatment planning, and surgical navigation. Although deep learning models have achieved remarkable success in this area, their strong reliance on large-scale datasets with fully segmented annotations remains a major obstacle to their practical application in real-world clinical environments. This dissertation aims to reduce the dependence on costly and labor-intensive annotations by addressing three distinct yet closely related segmentation problems. Each problem is characterized by a different level and type of supervision employed during model training and inference. In the first problem, weakly supervised segmentation, the model is trained using 3D medical volumes in which only a single 2D slice per volume is annotated. This problem serves as an initial step toward reducing labeling costs and evaluating the model’s ability to generalize from sparse supervision. The second problem, self-supervised few-shot segmentation, involves training entirely on unlabeled data, and performing segmentation during inference using only a single support volume that contains a few labeled slices. This scenario is more practical and cost-effective than the first, requiring minimal human annotation only at test time. In the third problem, inference-time segmentation, even inference relies on only one annotated slice from the target volume, which is used to propagate labels across the remaining unlabeled slices. Here, the model is trained without any supervision, and a minimal user-guided annotation is applied only during inference. Together, these three problems form a coherent continuum of progressive supervision reduction, moving toward the development of segmentation models that are more scalable, less dependent on annotations, and better suited for real-world clinical applications. The results demonstrate that even with minimal slice-level supervision, it is possible to develop models that achieve competitive performance compared to fully supervised methods.
  9. Keywords:
  10. Semantic Segmentation ; Few-Shot Three Dimentional Medical Image Segmentation ; Deep Learning ; Weakly Supervised Learning ; Self-Supervised Learning ; Few-Shot Learning ; Semi-Supervised Learning ; Medical Images Segmentation ; Mask Propagation

 Digital Object List

 Bookmark

...see more