Detective work.

Top Line: Who doesn’t look forward to figuring out what’s new and what’s not on a surveillance MRI following treatment of multiple brain metastases?

The Study: There may be a few of you weirdos out there, but for the rest of us there is emerging AI. This ancillary imaging study of NRG CC001 assessed various deep learning models to determine the most accurate one for brain metastasis detection and delineation on MRIs used in the real world. The good news for reproducibility is this dataset contained standard contrasted T1 and T2 FLAIR MRI series acquired at 98 different institutions all with their own pulse sequences and acquisition parameters done with the actual intent of detecting new or worsening brain mets. The winning model on the validation cohort (326 patients from 78 institutions) had a sensitivity of 86% with roughly 2 false positives per scan. This was then validated in a second testing cohort  (81 patients from 20 institutions) where it proved to have a sensitivity of 91% with an average of 1.7 false positives per scan, and it detected 7 additional brain mets (for a total of 327 detected) missed during manual contouring. For brain mets over 1.2 cm (72% of all mets in this dataset), sensitivity was 98% with an average of only 0.3 false positives per scan, meaning you could calibrate the detection size threshold to whatever you find most helpful / least annoying. Sure, the process still requires physician eyes, but such software can do a lot to keep those eyes from crossing.

TBL: Auto-detection of brain mets on initial diagnostic and post-radiation surveillance imaging of brain mets can improve clinical efficiency and physician detection rates. | Liang, Int J Radiat Oncol Biol Phys 2022