Harrison.Rad 1.5 claims a milestone in medical AI, but its benchmark comes with a significant asterisk.
Harrison AI built HR1.5 through a three-stage training process: domain adaptation on radiology reports, contrastive vision-encoder training on roughly 6 million image-report pairs, and fine-tuning on multi-turn clinical conversations. The model handles interleaved text and images across plain-film radiology types — chest, musculoskeletal, abdominal, spine, pelvic, and mammography. Harrison evaluated it against RadBench, a simulated version of the FRCR 2B Short Case exam scored against Angoff-method thresholds, as well as ReXGradient and internal datasets. Among the systems included in that evaluation, HR1.5 was the only one to clear the simulated passing standard and led accuracy on closed-format clinical questions across anatomical regions.
The underlying problem is real: imaging demand is growing faster than the radiology workforce can expand, and no recruiting push will close the backlog quickly. A model that can draft structured reports from images, clinical history, and prior studies could meaningfully reduce how long radiologists spend on routine documentation.
The asterisk: Harrison selected which competing systems to include, benchmarked against a simulated exam rather than an actual licensing threshold, and the paper frames this explicitly as groundwork for future clinical evaluation — being the best on a self-defined leaderboard is not the same as being ready for a hospital.