RADAR learns 146 abdominal CT findings from the reports radiologists already wrote

Report supervision is what makes the breadth affordable. RADAR splits a CT volume into anatomical regions and associates each region with the descriptions of it in the radiology report ^[10]. Across the training set that comes to more than 35 anatomy-aware image-text pairs per examination ^[21]. No clinician had to draw or label every example by hand, disease by disease ^[9]. Jianpeng Zhang, the senior algorithm researcher at DAMO Academy on the team, works on medical-image analysis, multimodal learning and vision-language pretraining ^[19].

The supervision is therefore whatever the reporting radiologists wrote. For the reported numbers to transfer, a hospital's scans have to be contrast-enhanced, because that is what the model reads ^[2], and its reporting vocabulary has to carve the abdomen up roughly the way the training reports did. Runtimewire's report says patient populations, scanners, reporting standards and disease prevalence in routine workflows can all shift the balance between missed findings and false alarms ^[18].

AUC measures how reliably a model separates positive from negative cases across thresholds, and it is not the percentage of patients diagnosed correctly ^[12]. A deployment picks thresholds, one per finding, and a mean over 146 findings does not say where any of them should go.

An AAAS summary of the study reports 0.895 across cohorts from eight external medical centers and 0.904 across more than 27,000 emergency CT cases ^[14]. Those sit 0.018 and 0.009 below the study's overall mean ^[23]. The same summary puts the best competing vision-language model in the researchers' evaluation at 0.776, which is 0.137 below RADAR ^[13][24].

The Science paper reports RADAR's average accuracy above that of 23 of the 26 radiologists it was compared against, drawn from multiple hospitals ^[15]. Three were at or above it ^[22]. The report says that comparison supports a second-reader or worklist role, and does not establish that the model can operate independently or improve patient outcomes ^[26].

The count in circulation needs care. The South China Morning Post described the model as detecting cancer and nearly 150 conditions ^[5]. The study evaluates 146 radiological findings, a category that includes cancers alongside other diseases and abnormalities ^[6], and runtimewire's report calls RADAR a broad abdominal imaging model, not a detector for 146 different cancers ^[7].

The repository is Apache 2.0 ^[3], which settles reuse and nothing about clearance. The team's phrase for RADAR, the world's first "expert-level generalist medical imaging model", is its own characterization and not a regulatory designation ^[17]. The same report says the breadth could shorten development cycles while regulation and prospective clinical evidence remain the gatekeepers ^[25].