Removing segmentation inconsistencies with semi-supervised non-adjacency constraint.

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Deep learning advances medical image analysis, yet segmentation faces challenges like data variability and exploiting anatomical knowledge.
• Accurate segmentation of anatomical regions is crucial for diagnosis and treatment planning.

Purpose of the Study:

• To improve region-labeling consistency in medical image segmentation using deep learning.
• To address the challenge of incorporating anatomical knowledge into segmentation networks.

Main Methods:

• Introduction of NonAdjLoss, an adjacency-graph based auxiliary training loss.
• NonAdjLoss penalizes anatomically-incorrect adjacency relationships in segmentation outputs.
• The method supports both fully-supervised and semi-supervised training extensions.

Main Results:

• Substantial reduction in segmentation anomalies observed.
• Demonstrated effectiveness on brain MRI (MICCAI-2012, IBSRv2) and whole-body CT (Anatomy3) datasets.
• Performance significantly improved with the semi-supervised training extension.

Conclusions:

• NonAdjLoss enhances deep learning-based medical image segmentation by enforcing anatomical consistency.
• The proposed loss function is effective in reducing segmentation errors, particularly in semi-supervised settings.
• This approach offers a promising direction for more reliable automated medical image analysis.