Room: Exhibit Hall | Forum 5
Purpose: To clinical assessment a deep learning-based auto segmentation algorithm for OARs segmentation in NPC CT images.
Methods: Three-hundred-sixty-four patients with nasopharyngeal carcinoma were used to establish a deep learning-based auto segmentation model in this study. Fifty ROIs (including spinal cord, brain stem, left and right temporal lobe, left and right eye, left and right optic nerve, left and right lens, left and right parotid, oral cavity, larynx, and chiasm) was segmented on each slice of CT images by various experienced radiation oncologists and all a second oncologist double-checked all contours. A 2D U-net structure was used and optimized during model training. Another 30 patients with clinical approved segmentation were used for performance evaluation. Two types of assessment were performed, including physician evaluation and objective evaluation. For physician evaluation, auto and manual segmented ROIs were mixed and anonymized. Physician cannot figure out ROIs type base on its name or other information. We asked two physicians, who have 5-yearsâ€™ experience, to selected the ROIs they prefered. For objective evaluation, the indices included Hausdorff distance (HD), average surface distance (ASD), Dice index (DSC), and Jaccard index (JSC).
Results: Based on the blinded assessment, physicians preferred all auto segmentation for all ROIs except brain stem and larynx. They have highly preferred auto segmentation for oral cavity and chiasm. For objective assessment, all ROIs DSC were around 0.8 except optic nerves (about 0.6) and oral cavity (about 0.7).
Conclusion: The physician cannot distinguish between manual segmentation and automatic segmentation. And for most ROIs, they more prefer auto segmentation.