Supervision by Denoising.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Learning-based image reconstruction models, like U-Nets, require extensive labeled data for generalization.
• Acquiring pixel/voxel-level labeled data is costly and challenging, especially in medical imaging due to inherent label variability.
• Traditional semi-supervised learning for image reconstruction often requires laborious, hand-crafted regularizers.

Purpose of the Study:

• To introduce a novel semi-supervised learning framework for image reconstruction that overcomes data scarcity.
• To develop a method that reduces the need for manual regularizer design in reconstruction tasks.
• To improve the generalization and accuracy of image reconstruction models using unlabeled data.

Main Methods:

• Propose "supervision by denoising" (SUD), a framework that uses a model's own denoised output as supervisory signals.
• Unify stochastic averaging and spatial denoising within a spatio-temporal denoising framework.
• Alternate denoising steps with model weight updates in an optimization process for semi-supervision.

Main Results:

• Demonstrate significant improvements in image reconstruction accuracy compared to supervised-only and ensembling methods.
• Successfully applied SUD to 3D anatomical brain reconstruction and 2D cortical parcellation.
• Validated the effectiveness of SUD in biomedical imaging applications with limited labeled data.

Conclusions:

• SUD offers an effective and less labor-intensive approach to semi-supervised learning for image reconstruction.
• The framework successfully leverages unlabeled data to enhance model performance in challenging imaging domains.
• SUD presents a promising direction for improving medical image reconstruction and analysis.