Impact of reconstruction kernel variability on segmentation consistency in low-dose thoracic CT

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Computed tomography (CT) segmentation is crucial for medical image analysis.
• CT reconstruction kernels (soft vs. hard) influence image characteristics.
• Variability in CT kernels may affect automatic segmentation accuracy and clinical decisions.

Purpose of the Study:

• To systematically investigate the impact of CT reconstruction kernel variability on thoracic segmentation using TotalSegmentator.
• To quantify segmentation consistency across different thoracic regions and scanner manufacturers.

Main Methods:

• Analysis of 9,529 paired CT scans from the National Lung Screening Trial (NLST) dataset, reconstructed with soft and hard kernels.
• Utilized TotalSegmentator for automatic segmentation of 84 thoracic regions.
• Quantified variability using Dice similarity coefficient, volumetric, and intensity differences.
• Assessed manufacturer-specific differences (Siemens, Philips, GE).

Main Results:

• Segmentation consistency varied significantly across thoracic regions; heart was more robust than the thyroid gland.
• Siemens scanners demonstrated the lowest segmentation variability between soft and hard kernels.
• Kernel-induced variability impacts automatic segmentation performance differently across anatomical regions.

Conclusions:

• CT reconstruction kernel choice influences thoracic segmentation consistency, with implications for clinical and research workflows.
• The findings underscore the need for standardized CT reconstruction protocols or kernel harmonization.
• Understanding kernel variability is essential for reliable automated medical image analysis.