A reciprocal learning strategy for semisupervised medical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semisupervised learning addresses the scarcity of annotated data in medical image segmentation.
• High-performance 3D convolutional neural networks (CNNs) require abundant segmentation masks, which are difficult to obtain.
• Existing methods often generate pseudo-labels without explicitly evaluating their quality.

Purpose of the Study:

• To introduce a reciprocal learning strategy for improved semisupervised volumetric medical image segmentation.
• To generate more reliable pseudo-labels for unannotated medical imaging data.
• To overcome limitations in current semisupervised segmentation by explicitly evaluating pseudo-label quality.

Main Methods:

• A reciprocal learning strategy employing a teacher-student network architecture.
• The student network learns from pseudo-labels generated by the teacher network.
• The teacher network optimizes parameters based on student performance feedback on annotated images, evaluated on pancreas CT, left atrium MR, and breast MR datasets.

Main Results:

• The proposed method achieved high performance across three datasets, with average Dice scores of 84.77% (pancreas), 90.46% (left atrium), and 78.53% (breast).
• Achieved Jaccard indices of 73.71%, 82.67%, and 69.00%, respectively, outperforming several state-of-the-art semisupervised methods.
• Demonstrated feasibility for challenging semisupervised segmentation tasks using only 20% labeled data.

Conclusions:

• The reciprocal learning strategy offers a general solution for semisupervised learning in medical image segmentation.
• This approach shows potential for application in other 3D segmentation tasks.
• The method effectively improves segmentation accuracy with limited annotated data.