Bone segmentation in contrast enhanced whole-body computed tomography.

Area of Science:

• Medical Imaging
• Radiology
• Computer Vision

Background:

• Accurate segmentation of bone regions in CT imaging is crucial for diagnostics, disease characterization, and treatment monitoring.
• Low-dose whole-body CT protocols reduce image quality, complicating the segmentation of contrast-enhanced regions due to difficulties in separating pixel intensities.

Purpose of the Study:

• To develop and validate a U-net architecture with novel preprocessing techniques for accurate segmentation of bone and bone marrow regions.
• To address the challenges of segmenting bone from contrast-enhanced regions in low-dose whole-body CT scans.

Main Methods:

• Implementation of a U-net architecture.
• Novel preprocessing techniques including windowing of training data.
• Modification of sigmoid activation threshold selection for improved differentiation.

Main Results:

• Achieved mean Dice coefficients of 0.979 ± 0.02 on internal dataset 1.
• Achieved mean Dice coefficients of 0.965 ± 0.03 on internal dataset 2.
• Achieved mean Dice coefficients of 0.934 ± 0.06 on an external test dataset.

Conclusions:

• Appropriate preprocessing is critical for differentiating bone from contrast dye in CT imaging.
• The proposed method demonstrates excellent segmentation performance even with limited data.
• The U-net architecture with novel preprocessing offers a robust solution for segmenting bone regions in challenging low-dose CT scans.