Head and neck automatic multi-organ segmentation on Dual-Energy Computed Tomography

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence

Background:

• Deep learning-based automatic segmentation is crucial for delineating organs-at-risk in radiation oncology.
• Dual-energy CT (DECT) offers enhanced contrast imaging, potentially improving manual and automatic delineation.
• This study evaluates a commercial auto-segmentation software's performance on DECT-generated images.

Purpose of the Study:

• To assess the performance of a commercial auto-segmentation software using various DECT image reconstructions.
• To compare auto-segmentation results against ground truth delineations for head and neck structures.
• To identify specific organs where DECT acquisitions are most beneficial for auto-contouring.

Main Methods:

• Seventy-four head and neck (HN) patient datasets with different DECT image types were analyzed: PEI80-DD, PEI80, PEI120, and VMI40.
• Auto-segmentations were compared to expert-defined ground truth (GT) delineations.
• Performance metrics included Dice Similarity Coefficient (DSC), 95th percentile Hausdorff distance (95HD), and mean surface distance (MSD).
• Qualitative evaluation of thyroid, parotid, and lymph node delineations was performed.

Main Results:

• Auto-segmentation software showed poor performance on thyroid and lymph node level II with PEI80-DD and VMI40 images.
• Parotid gland auto-segmentations received excellent qualitative scores across all image types.
• PEI80-DD images achieved the highest DSC scores, significantly outperforming other reconstructions for all evaluated organs (p < 0.05).

Conclusions:

• The auto-contouring system demonstrates limited generalizability to images generated from DECT acquisitions.
• Identifying specific organs that benefit from DECT is crucial for adapting and improving training datasets.
• Further research is needed to optimize auto-segmentation performance with DECT data for radiation oncology applications.