Investigating geometric and dosimetric accuracy of autosegmentation contours in stereotactic body radiation therapy for early peripheral non-small cell lung cancer

Background: the geometric and dosimetric accuracy of auto-segmentation oar are of key importance for radiation oncologists who use autosegmentation instead of manual segmentation. this study investigates the geometric and dosimetric accuracy of auto-segmentation oar for early peripheral nsclc using an artificial intelligence cloud online platform (ai contour). materials and methods: ai contour was used to perform the contour segmentation of oar on twenty patients with early peripheral nsclc, to evaluate geometric and dosimetric accuracies. manual segmentation and auto-segmentation were performed to depict the outlines of the heart, lung, trachea, esophagus, and spinal cord. for geometric accuracy, the authors acquired and compared the dice similarity coefficient, jaccard coefficient, hausdorf distance, center of mass deviation, inclusive index, and sensitivity index. for dosimetric accuracy, the dose statistical differences between manual- and auto-segmentation were analyzed. the absolute irradiation volume deviation (avd) and volume percentage deviation (vpd) for the v5, v10, v15, and v20 of the lungs were assessed. the absolute irradiation dose deviation (add) and dose percentage deviation (dpd) for oar were evaluated. results: the dsc for each oar was higher than 0.77. the dosimetric difference between manual and auto-segmentation was small and not significant (p>0.05). for the lung, the avd was less than 7 ml, the vpd was less than 3%, the add of oar was at most 0.4 gy, and the dpd was less than 4%. conclusion: the accuracy of the auto-segmented oar for early peripheral nsclc was acceptable based on ai contour.