Curriculum label distribution learning for imbalanced medical image segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Image Analysis

Background:

• Label distribution learning (LDL) shows promise for semantic segmentation by handling boundary ambiguity.
• Existing LDL methods struggle with imbalanced label distributions, where ambiguous regions are underrepresented, leading to biased learning and poor prediction of ambiguous pixels.

Purpose of the Study:

• To propose a novel curriculum label distribution learning (CLDL) framework to overcome the data imbalance issue in LDL-based semantic segmentation.
• To enhance the accurate prediction of ambiguous pixels by improving model learning on imbalanced datasets.

Main Methods:

• Developed a curriculum label distribution learning (CLDL) framework utilizing a task-oriented curriculum learning strategy.
• Introduced region label distribution learning (R-LDL) for balanced label distribution construction and improved imbalanced model learning.
• Proposed a task curriculum (TCL) for easy-to-hard learning by decomposing segmentation into multiple label distribution estimation tasks.
• Integrated a prior perceiving module (PPM) to connect learning stages using priors from easier stages.

Main Results:

• The CLDL framework effectively addresses imbalanced label distributions through balanced construction and prior perception.
• Experimental results on BRATS2018 and MM-WHS2017 datasets show significant improvements in segmentation metrics.
• The proposed method outperforms several state-of-the-art methods in semantic segmentation tasks.

Conclusions:

• The CLDL framework offers an effective solution for imbalanced label distribution learning in semantic segmentation.
• The combination of R-LDL, TCL, and PPM enhances model learning and segmentation accuracy, particularly for ambiguous regions.
• The study demonstrates the potential of curriculum learning strategies to improve LDL-based semantic segmentation performance.