RESEARCH WORK
Accepted: 2025-06-26
Objective To evaluate the accuracy of a deep learning-based intelligent contouring system in delineating head and neck organs at risk (OAR) and to explore its potential application in clinical radiotherapy. Methods Forty-five cases of head and neck tumor patients' positioning CT images were randomly selected. The SmartContour.AI software was used to contour 12 important OARs in the head and neck region (brainstem, spinal cord, left and right optic nerves, optic chiasm, left and right eyeballs, left and right lenses, left and right parotid glands, thyroid gland), and compared with the manual contouring by senior physicians. The consistency between automatic and manual contouring was quantitatively analyzed using seven geometric evaluation metrics, including Hausdorff distance (HD), Dice similarity coefficient (DSC), Jaccard coefficient, centroid deviation (CMD), sensitivity coefficient (SI), inclusiveness coefficient (II), and volume difference (VD). Additionally, the clinical acceptance of the intelligent contouring system was assessed by subjective evaluation from senior physicians. Results In terms of geometric accuracy, the maximum mean Hausdorff distance was 10.73 mm for the right optic nerve, and the minimum was 1.03 mm for the left lens. The mean DSC index for the OARs ranged from a minimum of 0.68 for the right optic nerve to a maximum of 0.95 for the right eyeball. The mean Jaccard coefficient for the OARs was highest for the right eye at 0.91 and lowest for the right optic nerve at 0.57. The maximum mean centroid deviation (CMD) was 4.08 mm for the right optic nerve and the minimum was 0.48 mm for the left lens. The sensitivity coefficient (SI) was relatively low for the left and right optic nerves and optic chiasm, at 0.63, 0.59, and 0.67, respectively, The SI value of the right eyeball is the largest, approaching 1.00. The mean inclusiveness coefficient (II) for all organs was ≥0.84. In terms of volume difference (VD), the brainstem, spinal cord, left and right eyeballs, left and right lenses, and thyroid gland were close to 0, while the VD values for the parotid gland, optic chiasm, and optic nerve exceed 0.10. In terms of clinical acceptance, the parotid glands, optic nerves, and optic chiasm had lower clinical acceptance, while the acceptance for the other organs was relatively high. Conclusion Automatic contouring of the brainstem, spinal cord, eyeballs, lenses, and thyroid gland showed high geometric consistency with manual contouring and was well accepted clinically, while automatic contouring of the parotid glands, optic nerves, and optic chiasm required further physician review and substantial modification.