μ-Net: Medical image segmentation using efficient and effective deep supervision.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep supervised learning methods have advanced medical image segmentation.
• Existing methods suffer from semantic differences in intermediate predictions and low learning efficiency.

Purpose of the Study:

• To propose novel deep supervised learning strategies (U-Net-Deep, U-Net-Auto) to address the semantic difference problem.
• To develop an efficient and effective deep supervised medical image segmentation model (μ-Net) by introducing a tied-weight decoder.
• To derive a Similarity Principle of Deep Supervision to guide future research.

Main Methods:

• Proposed U-Net-Deep and U-Net-Auto strategies to mitigate semantic differences.
• Introduced μ-Net with a tied-weight decoder for diverse pseudo-labels and faster convergence.
• Explored three μ-Net-based deep supervision strategies and derived the Similarity Principle of Deep Supervision.

Main Results:

• μ-Net significantly outperformed state-of-the-art baselines in both effectiveness and efficiency on four public benchmark datasets.
• Ablation studies validated the Similarity Principle of Deep Supervision, the tied-weight decoder, and the use of both segmentation and reconstruction pseudo-labels.

Conclusions:

• μ-Net offers an effective and efficient solution for deep supervised medical image segmentation.
• The Similarity Principle of Deep Supervision provides valuable guidance for future deep supervision research.