PLN: Parasitic-Like Network for Barely Supervised Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• 3D medical image segmentation is crucial but requires extensive manual annotation, which is costly and time-consuming.
• Existing methods struggle with the high cost of annotation for 3D datasets.
• Leveraging inter-slice and inter-volume similarities can reduce annotation burden.

Purpose of the Study:

• To develop a novel barely-supervised segmentation framework for 3D medical images.
• To investigate an extremely sparse annotation strategy using only one slice per 3D image.
• To reduce the labor, time, and expense associated with 3D medical image annotation.

Main Methods:

• Introduced a parasitic-like network comprising a registration module (host) and a semi-supervised segmentation module (parasite).
• Developed a parasitism mechanism with three stages (infection, development, eclosion) for module collaboration.
• Enabled inter-slice label propagation and inter-volume segmentation prediction using sparse labels.

Main Results:

• Achieved high performance on extremely sparse annotation tasks.
• Demonstrated effectiveness with Dice score of 84.83% on the LA dataset using only 16 labeled slices.
• Successfully generated accurate pseudo-labels for training through the proposed mechanism.

Conclusions:

• The proposed framework significantly reduces annotation requirements for 3D medical image segmentation.
• The parasitic-like network effectively couples delineation and model architecture for efficient segmentation.
• This approach offers a viable solution for accurate segmentation with minimal expert input.