Chest X-ray AI: detection, reporting and the VLM shift

On a chest radiograph, Milvue's AI flags several pleuro-pulmonary findings in a single pass, including pneumothorax, pleural effusion and airspace disease (alveolar opacities). These three targets were assessed in a multivendor study published in Radiology (2023, 2,040 radiographs), Milvue among the tools tested. The analysis is part of the Milvue Suite's Chest solutions, which cover 7 pathology families in one exam.

Yes, and it is one of the most studied use cases. In a comparison of commercial solutions in Radiology (2024) on lung-nodule detection, Milvue's AI exceeded the average of 17 radiologists (AUC 0.86 versus 0.81, p=0.04). A good score does not erase the difficulty: on a radiograph, a nodule remains an elusive finding, sometimes hidden by overlying ribs. The AI is a second look, never a verdict; the radiologist decides.

Performance is documented by peer-reviewed studies, not marketing figures. A review in Diagnostics (MDPI, 2023) sets out the opportunities and challenges of CE-marked chest X-ray AI; and the Radiology 2023 study is candid about the downside: on opacities, the tools produce more false positives than the radiologist's report. The chest radiograph remains a low-contrast exam, sensitive to the detector. That honesty about scope, more than a single score, is what clinical reliability rests on.

This is the 2026 shift: from detection to report drafting, driven by VLMs (vision-language models) that read the image and propose draft text. Milvue is building a radiology-native VLM, cited in Microsoft's June 2026 Community Hub blog for extending image-based reporting to musculoskeletal pathologies. Chest is the most heavily worked area on the market; Milvue also covers the other side of the worklist: musculoskeletal imaging, where its fracture-detection accuracy is the highest of the compared solutions. The model proposes a draft for review. The radiologist validates every line; it is not autonomous generation, and it is not a medical device. Our dedicated piece on the radiology VLM walks through this shift.

Results come quickly: a pre-report is available in under 30 seconds from the image, and the full analysis is returned to the PACS in under 2 minutes. This turnaround supports triage in the emergency department. Integration relies on DICOM and HL7 standards, and the detection solution is CE class IIa marked (EU MDR 2017/745).

No. The tool prioritizes and flags pleuro-pulmonary findings, never drawing the conclusion for the clinician. This decision-support stance runs through all of Milvue's detection, deployed at 600+ sites across 25 countries. It is the line that never moves: Our AI does not replace radiologists. It makes them irreplaceable.