Label tree semantic losses for rich multi-class medical image segmentation.

Area of Science:

• Artificial Intelligence in Medicine
• Medical Image Analysis
• Computer Vision

Background:

• Current AI segmentation models treat all errors equally, failing to utilize semantic relationships between labels.
• This limitation is amplified with complex datasets featuring numerous, subtly distinct classes.
• Existing methods struggle with efficiently segmenting medical images, especially with limited annotations.

Purpose of the Study:

• To develop and evaluate novel tree-based semantic loss functions for medical image segmentation.
• To enhance the exploitation of hierarchical label structures in AI models.
• To adapt these losses for training with sparse, background-free annotations.

Main Methods:

• Proposed two tree-based semantic loss functions that leverage hierarchical label organization.
• Integrated proposed losses with a recent approach for training using sparse, background-free annotations.
• Conducted experiments on head MRI (whole brain parcellation) and neurosurgical hyperspectral imaging (scene understanding).

Main Results:

• Demonstrated consistent improvements over task-specific baselines in both segmentation tasks.
• The Wasserstein-based compound loss showed strong performance in whole-brain parcellation.
• Hierarchy-weighted top-level supervision proved effective for sparse hyperspectral imaging (HSI).

Conclusions:

• Tree-based semantic loss functions effectively utilize hierarchical label information for improved medical image segmentation.
• The proposed methods offer enhanced performance in both fully supervised and sparsely annotated scenarios.
• These advancements hold significant potential for AI-defined clinical practice, including surgical planning and neuroimaging.