Deep Hierarchy-Aware Segmentation: A Novel Framework for MRIs Brain Tumor Segmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Effective brain tumor segmentation relies on exploiting label hierarchy.
• Current methods struggle with hierarchical prediction dependency and label similarity, limiting interpretability and accuracy.

Purpose of the Study:

• To present a novel framework, deep hierarchy-aware segmentation (DHAS), for interpretable and accurate brain tumor segmentation.
• To address limitations in existing methods regarding hierarchical dependency and label similarity.

Main Methods:

• DHAS generates hierarchical predictions by outputting pixel-wise probability conditional on parent labels, trained from conditional to unconditional probability.
• A tree-triplet loss is proposed to utilize label similarity by imposing hierarchy-induced distance in the feature embedding space.

Main Results:

• DHAS demonstrated significantly superior performance on BraTS2018, BraTS2019, and BraTS2020 datasets compared to other hierarchy-exploiting methods.
• The framework achieved a top 5 rank among 383 participants in the Brats2020 Challenge.
• Generalization was shown for cardiac segmentation on the ACDC dataset.

Conclusions:

• DHAS offers both hierarchically interpretable and high-accuracy predictions for brain tumor segmentation.
• The framework shows potential for clinical applications due to improved performance and interpretable outputs.
• The proposed methods are effective and generalize to other segmentation tasks.