Top-k Bottom All but σ Loss Strategy for Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Medical image segmentation is crucial but often hindered by noisy pixel-level annotations and limited data.
• Existing methods struggle with these challenges, impacting segmentation accuracy.

Purpose of the Study:

• To introduce a novel loss function envelope derived from the Top-k loss strategy for improved medical image segmentation.
• To address challenges of noisy pixel-level annotations and limited data in semantic segmentation tasks.

Main Methods:

• Developed a new loss function envelope utilizing Top-k strategy at image level and 'Bottom all but σ' strategy at pixel level.
• Implemented a derivative smoothing procedure to handle discontinuities in automatic learning differentials.
• Tested the approach with various backbone models including Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs).

Main Results:

• Achieved performance improvements in segmenting burned skin areas, fetal abdominal structures in ultrasound, and cardiac structures in MRI.
• Demonstrated successful application across diverse medical imaging modalities and anatomical regions.
• Validated the method's effectiveness on challenging segmentation tasks with degraded annotation quality.

Conclusions:

• The proposed mechanism enhances model training by selectively emphasizing loss values through complementary strategies.
• The approach is particularly beneficial in difficult segmentation scenarios with noisy or inconsistent pixel-level annotations.
• The method shows consistent performance across both CNN and ViT architectures.