Advancing Medical Image Segmentation with Transformers: Investigations into Self-Supervised and Supervised Learning
- Yan, Xiangyi
- Advisor(s): Xie, Xiaohui XX
Abstract
Medical image segmentation is crucial in medical imaging, aiding detailed anatomical assessments and supporting diagnostic and therapeutic decisions. The introduction of deep learning has revolutionized this field, improving the accuracy and efficiency of segmenting complex images like CT scans, MRIs, and ultrasound. Deep learning excels by extracting hierarchical features from vast data volumes, enabling precise structure delineation from organ to cellular levels.
Supervised learning dominates this space, requiring extensive annotated datasets where each image pixel is labeled with a class (e.g., tissue types). Convolutional Neural Networks (CNNs) are widely used due to their ability to handle spatial hierarchies and preserve local context, crucial for precise segmentation.
However, the scarcity and cost of expert-labeled images have catalyzed interest in selfsupervised learning, which learns from unlabeled data, diminishing the reliance on annotated datasets. This involves pre-training models on large unlabeled datasets with auto-generated labels and fine-tuning on smaller labeled sets. This method has improved model robustness and adaptability, particularly valuable in varied imaging and demographic conditions.