Microsoft Research splits chest X-ray AI between a writing model and a measurement tool

Chest X-rays are asked to do many different jobs in the same sitting: draft a radiology report, answer a clinician's question about whether a finding is present, locate that finding on the image, and produce numbers that downstream rules and care pathways can read. A clinically useful AI has to handle that range without trading fluency for accuracy. Microsoft Research is testing a split: one model writes the report, and a deterministic tool does the measuring. The writing half is CARE-X, a research model with no regulatory clearance. The workflow references in the blog post describe future research, not current capability.

CARE-X is built as a single vision-language model that can both generate text and produce structured predictions. The encoder is SigLIP2-so400M, a 400-million-parameter image backbone. The language side is Phi-4-mini-instruct at 3.8 billion parameters. A lightweight adapter connects them. On top, the team co-trains task-specific auxiliary heads for classification and grounding that share the same language backbone. Training runs in three supervised-fine-tuning stages: vision pre-training, adapter and head training, and a LoRA pass over the language model. A DAPO reinforcement-learning stage then optimizes task-specific rewards for clinical reporting, diagnostic accuracy, and spatial grounding. The result is one model that can be queried for a free-text report, a calibrated abnormality score, or a bounding box.

The numbers Microsoft reports land on the high end of every public benchmark the team cites. CARE-X posts the highest CRIMSON scores across MIMIC-CXR, IU-Xray, CheXpert-Plus, and ReXGradient, hits a 0.913 F1 on the Chest ImaGenome abnormality classification task against CheXOne's 0.866 and MedGemma's 0.839, and climbs to 94% overall accuracy on ReXVQA, a 41,007-pair question-answering benchmark where it ranks first on the ReXrank leaderboard as of August 2026, roughly six percentage points above the next-best publicly reported model. A small but telling detail: the DAPO reinforcement-learning stage closes the gap to a dedicated detection head on anatomy grounding, pushing CARE-X's generative output to 0.868 mAP versus the SFT auxiliary detector's 0.865 mAP. All of these are vendor-reported numbers on public benchmarks. None has been independently reproduced outside Microsoft, and the team is explicit that the model has not been cleared or approved for clinical use.

The measurement half lives in a separate experiment. The team pairs Qwen3-VL-4B-Instruct with deterministic measurement tools. The vision-language model identifies a structure on the image and emits a tool call. A rule-based function then reads pixel coordinates and returns the exact number. The result is a 43.6-percentage-point average lift in F1 on measurement-dependent conditions, including cardio-thoracic ratio and aortic measurements. Ascending Aorta Enlargement goes from 39.33 to 100.00; Descending Aorta Enlargement from 28.57 to 100.00. The division of labor is the point: the VLM's job stops at perception, and the tools compute exactly. This is a research result on Qwen3-VL-4B, not a CARE-X number, and conflating the two would be a misread.

The validation cohort gives the work its real-world weight. Microsoft partnered with Narayana Health on a retrospective, ethics-approved, de-identified evaluation of 1,047 chest radiographs across five rare ICU conditions with prevalences between 2.6% and 5.2%, plus 122 CT-confirmed positive enlargement cases. The tool-augmented variant reached 94.26% recall on the CT-confirmed cohort, 10.65 percentage points above the best perception-only baseline. In a related EACTS 2026-accepted study, the same architectural pattern surfaced mild aortic dilation in 40 of 43 CT-confirmed cases (93% sensitivity), against 5 of 43 (12%) on the initial radiology reads, surfacing 35 additional cases the original workflow missed. The model is a finalist for the IHF Innovation Hub at the World Hospital Congress 2026, a recognition signal separate from the benchmark claims above.

The data does not yet support a clinical story. The Narayana cohort is single-site, retrospective, and Indian; generalization to other populations, equipment, and protocols is unproven. The enlargement analysis reports recall only. Specificity and positive predictive value are not yet reported, and the team flags an extended study as forthcoming. None of the benchmark numbers has been independently reproduced. And the architectural bet, a combination of unified generation, discriminative heads, reinforcement-learning reward, and tool calls for exact measurement, is one path, not a settled one. If a clinically cleared radiology VLM ships from a competitor using a different architecture, a leaderboard result alone will not have vindicated this approach.

The next read is the extended Narayana study when it lands. Recall, specificity, and positive predictive value on the same CT-confirmed cohort, across more than one site, would either tighten the case or expose the limit.