Integrating orthogonal supervision for sparse semi-supervised 3D medical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semi-supervised learning (SSL) advances 3D medical image segmentation.
• Fully annotating volumetric data is time-consuming and costly.
• Sparsely annotated SSL uses limited labeled data but often yields suboptimal performance due to view-specific training.

Purpose of the Study:

• To develop a more effective semi-supervised learning framework for 3D medical image segmentation using sparse annotations.
• To address the limitations of view-specific training in current sparsely annotated SSL methods.
• To propose a novel strategy that integrates supervisory signals from all orthogonal views.

Main Methods:

• Proposed Integrating Orthogonal Supervision (IOS), a strategy using a single 3D encoder to learn from axial, sagittal, and coronal planes.
• Introduced a Tri-Decoder Framework for learning from both sparsely labeled and unlabeled data.
• Implemented a cross-decoder supervision strategy for unlabeled slices within labeled volumes and used joint pseudo-labels for unlabeled volumes.

Main Results:

• The IOS framework demonstrated superior performance compared to state-of-the-art methods on LA, Pancreas, and KiTS19 datasets under sparse supervision.
• Simultaneous integration of orthogonal views led to a richer understanding of volumetric data distributions.
• The proposed Tri-Decoder Framework and cross-decoder supervision improved the quality of pseudo-labels and overall segmentation accuracy.

Conclusions:

• Integrating orthogonal views within a single framework is more effective for sparsely annotated 3D medical image segmentation.
• The proposed IOS strategy with its Tri-Decoder Framework offers a robust solution for leveraging limited annotations.
• This approach significantly improves 3D medical image segmentation accuracy, offering a more efficient alternative to full annotation.