RADT-10. THE LOST METASTASES: DEEP LEARNING’S POTENTIAL IN RADIOSURGERY QUALITY ASSURANCE

Abstract Introduction Identifying, segmenting, measuring, and following multiple brain metastases treated with radiosurgery can be time consuming error prone. Machine learning has shown promise for automated detection segmentation. Recently, a U-Net inspired model combining volume aware loss functions sampling methods was trained in an industrial-academic partnership. A total of 530 clinically annotated T1 gadolinium MRIs were used. Initial validation showed high sensitivity (91%) average 0.66 false positives per MRI. The goal the present work to characterize those “false positives” which may represent undetected metastases. METHODS images used development planning. Lesions had first been identified by radiologist, second clinicians during tumor board review, third treating radiation oncologist neurosurgeon (potentially after segmentation trainee) finally fellow oncologists quality assurance rounds. Despite these checks, 10 patients (2%) lesions considered potential clinical misses when all manually reviewed single investigator. Further detailed review including prior subsequent imaging arbitrate nature lesions. RESULTS Among cases, four confirmed as metastases: two required 2 died further imaging. six other adjudicated true (typically vascular). CONCLUSION multi-tier workflow at our institution left very few unidentified (0.8%). this low rate, AI algorithm still detected that treatment. Future investigations will focus on roles simplifying accelerating workflow. It also remains established if more would seen community settings where workflows include fewer sequential reviews.