Rethinking boundary detection in deep learning-based medical image segmentation.

Area of Science:

• Computer Vision
• Medical Image Analysis

Background:

• Accurate medical image segmentation is crucial but struggles with precise boundary delineation.
• Existing methods often face challenges in segmenting complex boundary regions effectively.

Purpose of the Study:

• To introduce a novel network architecture, CTO, for enhanced medical image segmentation, particularly focusing on boundary accuracy.
• To improve the balance between segmentation accuracy and computational efficiency.

Main Methods:

• CTO employs a dual-stream encoder combining Convolutional Neural Networks (CNNs) for local features and Vision Transformer (ViT) for long-range dependencies.
• A boundary-guided decoder utilizes explicit edge detection operators to enhance learning of boundary areas.
• The architecture follows an encoder-decoder paradigm, integrating CNNs, ViT, and edge detection.

Main Results:

• CTO achieved state-of-the-art segmentation accuracy across seven diverse medical imaging datasets.
• The proposed method demonstrated a superior balance between accuracy and efficiency compared to existing techniques.
• No additional data or label injections were required for CTO's performance enhancement.

Conclusions:

• CTO offers a significant advancement in medical image segmentation, particularly for challenging boundary regions.
• The hybrid CNN-ViT architecture with boundary guidance provides a robust and efficient solution.
• CTO sets a new benchmark for accuracy and efficiency in medical image segmentation tasks.