Ipnet: informative patches learning for semi-supervised magnetic resonance image segmentation

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Medical image segmentation is crucial but hindered by high labeling costs.
• Magnetic resonance images present challenges like low contrast and variable organ appearance.
• Existing semi-supervised methods struggle with these complex image characteristics.

Purpose of the Study:

• To develop an advanced semi-supervised method for segmenting magnetic resonance images.
• To specifically target and improve the learning of challenging image regions.
• To enhance segmentation accuracy in the presence of low contrast and anatomical variability.

Main Methods:

• Proposed IPNet (Informative Patches Network) utilizes a novel scoring strategy based on prediction uncertainty and category diversity.
• Implemented a patch-swapping technique to create new training samples from informative patches.
• Introduced global and local consistency losses to refine learning on these augmented samples.

Main Results:

• Experiments on ACDC, PROMISE 12, and LA datasets demonstrated significant improvements.
• The proposed method effectively handles challenging magnetic resonance images.
• IPNet achieved superior performance compared to existing semi-supervised segmentation techniques.

Conclusions:

• IPNet offers a robust solution for semi-supervised medical image segmentation.
• The informative patch learning strategy is effective for improving segmentation in challenging datasets.
• This method shows great potential for clinical applications requiring accurate medical image segmentation.