Semi-Supervised Medical Image Segmentation Based on Frequency Domain Aware Stable Consistency Regularization.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Semi-supervised learning is crucial for medical image segmentation due to limited annotated data.
• Consistency regularization methods improve training but often use artificial perturbations, introducing bias.
• Existing methods neglect encoder-stage supervision and spatial-frequency information, hindering initial learning and overall performance.

Purpose of the Study:

• To propose a novel semi-supervised medical image segmentation approach.
• To address limitations of artificial perturbations and lack of encoder supervision in current methods.
• To leverage inherent spatial-frequency information for more stable and effective model training.

Main Methods:

• Developed a frequency domain aware stable consistency regularization technique.
• Utilized inherent image frequency components (high and low) as consistency constraints, avoiding artificial perturbations.
• Incorporated supervision in the encoder stage to preserve feature space integrity during training.

Main Results:

• The proposed method effectively overcomes biases introduced by artificial perturbations.
• Encoder supervision ensures robust initial feature learning, preventing chaotic learning phases.
• Experimental validation confirms the effectiveness of the frequency domain aware approach for semi-supervised medical image segmentation.

Conclusions:

• Frequency domain aware stable consistency regularization offers a promising direction for semi-supervised medical image segmentation.
• Integrating encoder supervision and utilizing intrinsic image properties enhances model robustness and accuracy.
• This approach provides a more stable and less biased alternative to traditional consistency regularization methods.