Microsoft adds precise measurement tools to Qwen for radiology

The model stops just looking and starts calling functions

Microsoft has published the CARE-X research experiment, connecting the Qwen3-VL-4B-Instruct model with deterministic measurement tools. The model maintains visual access to a chest radiograph and can call an external function to identify anatomical landmarks or calculate distances. It then synthesizes a diagnosis only after combining its visual perception with an exact number from the tool.

Visual estimation is not enough for heart diseases

For radiology, this represents a fundamental difference in approaching threshold-dependent diagnoses. Aortic dilation or cardiomegaly are not evaluated by feeling, but by clear metrics. While previous vision-language models tried to estimate heart size directly from pixels, the new method allows the model to obtain hard data. Tool-augmented inference increased the F1 score for cardiomegaly by more than 21 points compared to purely visual estimation.

It is a long way from research to clinical reality

The inference-time pipeline runs as a research prototype and required no task-specific training. In a hospital, however, the model will encounter diverse equipment calibrations and less-than-ideally positioned patients. Furthermore, relying on an external measurement tool means that if the model sends wrong coordinates, the tool returns a precisely calculated nonsense. Quality assurance thus shifts from checking the text itself to checking the accuracy of function calls.

Localization accuracy will be the critical point

Future development will be determined by the ability of models to safely handle more complex coordinates within the body. Whether this architecture succeeds outside the laboratory will be shown by the success rate of locating key anatomical landmarks on images with unexpected pathologies.

Lilith's verdict

Doctors don't need a model that looks at an image and tries to guess if the heart is too big. They need an expert system that takes a ruler, measures it, and backs up its opinion with a number.