1334 Automated Tumor Segmentation for SRS: How Does Artifical Intelligence Stack up to Natural Intelligence?

INTRODUCTION: Radiosurgery is now the standard of care for multiple brain metastases, with no proven upperlimit of number of lesions that can be treated. Each individual brain metastasis has to be segmented by a combination of graphical interface tools, a process that can be rate limiting. We explored the automated tumor detection and segmentation algorithm degned by Brainlab Inc. inc comparison to the manual segmentation. METHODS: Automated lesion detection is a new software package from Bainlab Inc. which is fully automatic. Based on a Neural network and GPU it simultaneously delineates multiple cranial tumors on contrast-enhanced T1-weighted MRI scans. MRI scans obtained for targeting were simultaneously processed through the AI engine and manually reviewed and segmented by experienced users. At least two users reviewed each manual segmentation. RESULTS: In the Fifty patients with 342 tumors the median number of brain metastases was 4 (1-31). Over-all the AI engine detected 88% of all brain mets. Lesions missed by the engine had a mean volume of 0.95 cc and were located close to confounding structres in more than half the patients. False positive detections rejected on reviiew occured in 14 tumors. 11 tumors were missed by natural intelligence and were detected by AI. In two instances tumors close to each other were segmented as one. Time to segmentation: Manual segmantation took a mean time of 45 seconds per tumor. AI segmention of upto 31 lesions was complete in 65 seconds. Review for the case with 31 brain metastasis to accept the AI results took 5 minutes. CONCLUSIONS: AI based segmentation can be a valuable assitant in modern high lesion count high volume radiosurgery practice.