Translation Consistent Semi-supervised Segmentation for 3D Medical Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• 3D medical image segmentation is crucial but requires extensive annotated data, which is costly to acquire.
• Semi-supervised learning (SSL) reduces annotation burden by using both labeled and unlabeled data.
• Existing SSL methods based on consistency learning may focus on spatial context instead of object features.

Purpose of the Study:

• To develop a novel SSL method for 3D medical image segmentation that overcomes limitations of current approaches.
• To improve the model's ability to learn segmentation patterns directly from foreground objects.
• To enhance training convergence and robustness against pseudo-labeling errors in co-training.

Main Methods:

• Introduced Translation Consistent Co-training (TraCoCo), an SSL method that perturbs input data views by varying spatial context.
• Proposed a Confident Regional Cross entropy (CRC) loss function to improve training and robustness.
• Evaluated the method on multiple 3D and 2D medical image segmentation benchmarks.

Main Results:

• TraCoCo achieved state-of-the-art (SOTA) results on the Left Atrium (LA), Pancreas-CT (Pancreas), and BraTS19 benchmarks.
• The method also demonstrated superior performance on the 2D Automated Cardiac Diagnosis Challenge (ACDC) benchmark.
• The proposed CRC loss improved training convergence and robustness.

Conclusions:

• TraCoCo effectively learns segmentation patterns from foreground objects by varying spatial context, outperforming existing SSL methods.
• The combination of TraCoCo and CRC loss offers a robust and efficient solution for 3D medical image segmentation.
• The method's success across diverse benchmarks highlights its broad applicability and effectiveness.