The impact of neck tilt on the accuracy of deep learning generated contours for CT images of the head and neck.

BACKGROUND

Deep learning auto segmentation (DLAS) tools are widely used for radiation therapy planning, yet limited information exists on how patient positioning impacts their performance, particularly in head and neck (HN) computed tomography (CT) imaging.

PURPOSE: This study investigates the impact of abnormal neck tilt on the accuracy of DLAS-generated contours in HN CT scans. We hypothesize that abnormal positioning degrades auto segmentation performance, particularly for organs at risk (OARs) influenced by cervical spine orientation.

METHODS

A total of 35 HN CT scans were retrospectively analyzed. Neck tilt was quantified using principal component analysis of the spinal cord contours from C1 to C4, with patients stratified into normal and abnormal tilt groups based on percentile thresholds. Seven FDA-cleared DLAS tools were evaluated for OAR segmentation accuracy across brainstem, parotid glands, submandibular glands, brachial plexus, and optic nerves. Gold-standard contours were curated from manually created clinical contours and compared to DLAS output using Sorenson's volumetric dice similarity coefficient (DSC), surface Dice similarity coefficient with a 2 mm threshold (sDSC), mean distance to agreement (MDA) and difference in mean dose. Statistical differences were assessed using Wilcoxon rank-sum tests (p < 0.05).

RESULTS

Among the evaluated OARs, the parotid glands showed consistent and statistically significant degradation in contouring accuracy for patients with abnormal neck tilt across all metrics (DSC, sDSC, MDA) and across all seven DLAS tools. Changes in contouring accuracy resulted in significant differences in both absolute and signed difference in mean dose. No consistent differences were observed for other structures across multiple DLAS tools.

CONCLUSIONS

Abnormal neck tilt is associated with reduced DLAS accuracy and greater dose variability for parotid gland segmentation in HN radiotherapy planning. These findings underscore the need for enhanced patient quality assurance strategies for patients with abnormal positioning when using clinical DLAS workflows. Future work should explore automated detection of anatomical outliers and site-specific model retraining to ensure accurate contouring delineation for patients with atypical positioning or anatomy.