Loading…
Medical imaging vision-language model for chest X-ray, CT and MRI that generates reports, localizes lesions and compares studies over time.
No providers recorded yet. Browse all providers
A 25 mm lesion in the left frontal lobe seven years ago and a 45 mm lesion in the same place today are one disease observed twice, but a model that reads one image at a time sees two unrelated pictures. Radiology AI has largely been built as perception — name the finding, draw the box — while the judgement that follows the finding stayed with the clinician.
U2-RadiMed is a medical imaging multimodal model released by Unisound on 13 August 2026 that targets that gap. It is built on U2-Med, the company's "tri-medical" language model covering clinical care, medical insurance and pharmaceuticals, which itself sits on Unisound's general-purpose U2 base model; U2-RadiMed adds visual understanding on top of that clinical knowledge layer rather than training a radiology encoder in isolation. It spans chest X-ray, CT and MRI, and accepts mixed input — one image, several images, or images together with free text.
The design intent is a closed loop: detect the abnormality, retrieve the relevant medical knowledge, reason over it, and state a finding with the reasoning attached. That places it alongside medical vision-language models such as MedGemma and MedRegA, but with the distinguishing emphasis on longitudinal comparison and explicit differential diagnosis rather than single-study interpretation.
Unisound has not disclosed the architecture, parameter count, or the composition and scale of the training corpus, and there is no preprint or technical report. It has also not identified the vision encoder — coverage of the MedBench 5.0 release describes the model as a vision encoder trained jointly with the U2-Med text backbone, but the encoder's identity, size and pretraining are not stated. The comparative headline numbers are company-reported. On a comparative evaluation across medical imaging tasks, Unisound places U2-RadiMed first with an overall score of 55.47, ahead of Gemini 3.1 (47.9) and GPT-5.4 (47.5), and also compared against Hulu-Med-27B and Qwen3.6-27B; within the six-capability breakdown it reports top scores on four dimensions, including 37.58 on lesion detection and localization, 74.2 on general medical visual question answering, and 46.8 on advanced medical reasoning. The protocol and composition of that comparative suite are not published. The overall score of 57.2 on the MedBench 5.0 full-modality track is a separate and independent result: MedBench is a third-party leaderboard operated by OpenCompass (Shanghai AI Lab), and its 16 July 2026 iteration — a month before the launch announcement — placed U2-RadiMed first over 12 evaluation sets in three capability dimensions, with published per-dimension scores of 66.4 on medical visual perception and text extraction and 72.0 on clinical decision support and reasoning. Two worked cases accompany the release: a low-grade glioma followed across seven years of MRI, where the model tracked growth from 25 mm to 45 mm and read peri-resection high signal as expected post-operative change rather than residual tumor; and an orbital contrast CT where muscle-belly thickening with spared tendon insertions led it to thyroid-associated orbitopathy over idiopathic orbital inflammation, orbital lymphoma and vascular causes.
The intended setting is a hospital radiology department: screening studies for abnormalities, drafting structured reports, answering questions about a specific image region, tracking a tumor or a treatment response across serial imaging, and assembling a differential a clinician can check against the stated evidence. Unisound positions it for integration into imaging departments at hospitals of varying capability, where the reasoning trace matters as much as the prediction. Because distribution is API-only, use is limited to organizations working through the company's platform.
U2-RadiMed's interest lies in what it attempts rather than what has been independently verified: moving radiology models from labelling findings toward the clinical reasoning that surrounds them, and treating a patient's imaging history as a single sequence rather than a stack of unrelated inputs. Most of the evidence, however, is vendor-supplied: the comparative suite behind the 55.47 headline is vendor-run and unnamed, and there are no weights or code with which to reproduce any of it. The MedBench 5.0 placement is the exception — a third-party leaderboard result, though one that scores a general full-modality medical track rather than the longitudinal comparison and differential reasoning the model is pitched on. No regulatory clearance has been stated for the clinical decision support capability the model advertises, which in most jurisdictions is the gate between a research demonstration and use on patients.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.