Imaging: reports and image quality

In the 2022 clearance's reader study, three radiologists compared images reconstructed with and without the feature from the same data in 133 cases; processed images were equivalent or better for apparent signal-to-noise ratio and sharpness in all cases and preferred in 99% of evaluations. At New York University, 117 patients who had a 7-minute knee MRI with the feature and then arthroscopy were read by seven radiologists: sensitivity and specificity were 97% and 98% for anterior cruciate ligament tears and 89% and 87% for medial meniscus tears. In 50 volunteers in Rome, liver diffusion imaging took 207.3 seconds with the feature versus 513.6 seconds with a conventional high-resolution protocol.

In 74 healthy adults in Beijing, the feature improved image quality but shifted brain morphometry, raising hippocampal volumes by about 3.9%. A Nottingham comparison found other denoising methods gave better diffusion fiber orientation estimates.

At San Raffaele in Milan, three readers of differing experience read multiparametric MRI of 110 men, 102 of whom had biopsy; clinically significant cancer was found in 47%. With the software, the least experienced reader identified 127 lesions rather than 101 and that reader's negative predictive value rose from 0.65 to 0.82; the two more experienced readers did not change. At Sapienza University in Rome, four novice readers of 200 scans agreed with an expert more with the software for two readers and less for two (one fell from a kappa of 0.603 to 0.292). In 170 retrospective scans at the same center, the software did not significantly change diagnostic performance but shortened reporting time, from 8.23 to 7.11 minutes for an inexperienced reader.

The 2026 FDA summary reports a standalone case-level area under the curve of 0.819; 86% of high-suspicion lesions in its validation set were cancer. None of the three Italian studies listed company authors.

In the Moffitt study, 200 reports from four abdominal radiologists each had three impressions: the original, one drafted by Rad AI's model and one drafted by a general large language model, GPT-4.1. Independent radiologists showed no preference between Rad AI and original impressions (Cohen's h of 0.03, P=0.78); the original authors slightly preferred their own (h=0.18, P=0.07) and rated them more complete (2.87 vs. 2.74 on a 3-point scale). The general model's impressions averaged 75.1 words, against 41.2 for the originals and 34.2 for Rad AI, and were rated much less concise. Raters judged the likelihood of patient harm as low for all three.

Agreement between raters was poor to moderate (Krippendorff's alpha from −0.09 to 0.67). Seven of the 14 authors listed a Rad AI affiliation, including the corresponding author, who also listed Moffitt; no AI impression tool was in clinical use there during the study period.

In a prospective study at five institutions, 40 patients had standard brain MRI and volumetric sequences 60% shorter, processed with SubtleMR; two neuroradiologists rated the processed images better for perceived quality, and volumetric biomarkers and clinical classification did not differ. Authors were from RadNet, Subtle Medical, Cortechs.ai and Stanford. In the spine study at five centers, three neuroradiologists rated 40% faster processed scans better than standard ones for perceived signal-to-noise ratio and artifacts and equivalent for other features; a Subtle Medical employee was an author.

At Mount Sinai, 43 patients with knee pain had conventional (13 minutes 27 seconds) and accelerated processed (6 minutes 45 seconds) knee MRI; two radiologists found no significant differences in diagnosis, image quality or artifacts. One author was from Subtle Medical. A French study with no company authors halved brain acquisition time by halving the image matrix and used SubtleMR to restore resolution in 33 patients with metastases or meningiomas: specificity was 1 and lesion sensitivity 0.92 for radiologists and 0.77 for radiation oncologists; missed lesions were under 4 mm, and image quality was rated lower than on reference scans.