Improved distinct bone segmentation from upper-body CT using binary-prediction-enhanced multi-class inference.

Area of Science:

• Medical imaging
• Computer vision
• Biomedical engineering

Background:

• Automated distinct bone segmentation is crucial for surgical planning and navigation.
• Current 3D U-Net models struggle with differentiating bone tissue from background, limiting segmentation accuracy.

Purpose of the Study:

• To enhance multi-class distinct bone segmentation accuracy in CT scans.
• To address the primary error source in 3D U-Net segmentation: background vs. bone-tissue confusion.

Main Methods:

• Propose binary-prediction-enhanced multi-class (BEM) inference, incorporating a binary background/bone-tissue prediction.
• Evaluate BEM inference using a two-stage approach and networks with two segmentation heads.
• Test on in-house (16 upper-body CT scans) and public synthetic (50 CT scans) datasets.

Main Results:

• The most successful two-segmentation-head network achieved a class-median Dice coefficient of 0.85 on cross-validation for upper-body CT scans.
• BEM inference outperformed previous 3D U-Net baselines and reported results from other research groups.
• Improved segmentation results were also observed on the synthetic dataset using BEM inference.

Conclusions:

• Binary bone-tissue/background prediction guidance significantly improves distinct bone segmentation in CT scans.
• The BEM inference method demonstrates robustness across different approaches for obtaining binary predictions.
• The enhanced segmentation accuracy holds for both two-stage and two-headed network architectures.