SALT: Introducing a framework for hierarchical segmentations in medical imaging using label trees.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Traditional segmentation networks often overlook the hierarchical relationships between anatomical structures.
• This limitation hinders efficiency and interpretability in medical image analysis.
• Developing methods that incorporate anatomical hierarchies is crucial for advancing segmentation accuracy.

Purpose of the Study:

• Introduce Softmax for Arbitrary Label Trees (SALT), a novel activation function for medical image segmentation.
• Leverage hierarchical relationships in anatomical structures to improve segmentation efficiency and interpretability.
• Enable natural representation of anatomical hierarchies in CT imaging for enhanced segmentation.

Main Methods:

• Developed SALT, an activation function extending softmax to arbitrary hierarchical label trees by modeling conditional probabilities along parent-child relations.
• Trained and evaluated SALT on the SAROS dataset (900 scans, 113 labels), with validation and testing on subsets.
• Assessed performance using Dice scores across multiple datasets: SAROS, CT-ORG, FLARE22, LCTSC, LUNA16, and WORD, with 95% confidence intervals from 1000 bootstrapping rounds.

Main Results:

• SALT achieved top performance on LUNA16 (Dice 0.93) and SAROS (Dice 0.929).
• Demonstrated reliable accuracy on CT-ORG (0.891), FLARE22 (0.849), LCTSC (0.908), and WORD (0.844).
• Achieved rapid segmentation, processing a 1000-slice CT scan in an average of 35 seconds.

Conclusions:

• SALT effectively leverages hierarchical body structures for efficient and interpretable medical image segmentation.
• The method's speed and accuracy support integration into clinical workflows for automated whole-body segmentation.
• SALT has the potential to enhance diagnostic workflows and improve patient care through faster, more accurate image analysis.