An Annotation Sparsification Strategy for 3D Medical Image Segmentation via Representative Selection and...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep learning (DL) methods excel in medical image segmentation but require extensive data annotation.
• Voxel-wise labeling of 3D medical images is time-consuming and labor-intensive, posing a significant bottleneck.
• A performance gap exists between DL models trained on partially labeled versus fully annotated datasets.

Purpose of the Study:

• To develop a novel deep learning framework to reduce annotation effort in 3D medical image segmentation.
• To bridge the performance gap between full annotation and sparse annotation strategies.
• To minimize data redundancy and save annotation time while maintaining high segmentation accuracy.

Main Methods:

• Proposed a DL framework incorporating representative slice selection to minimize redundancy and annotation effort.
• Implemented a self-training strategy using pseudo-labels generated from models trained on selected annotated slices.
• Validated the framework on two public datasets: HVSMR 2016 Challenge and mouse piriform cortex.

Main Results:

• The proposed framework achieved competitive segmentation results compared to state-of-the-art DL methods.
• Demonstrated effectiveness using less than approximately 20% of the fully annotated data.
• Successfully reduced annotation effort while maintaining high performance in 3D medical image segmentation tasks.

Conclusions:

• The developed DL framework significantly reduces annotation requirements for 3D medical image segmentation.
• The combination of representative slice selection and self-training offers an effective solution to the data annotation bottleneck.
• This approach enables high-performance segmentation with substantially less labeled data, advancing medical AI applications.