Lab for Artificial Intelligence in Medicine

Radiograph Morphometry Agent Grounded reports for musculoskeletal radiographs

Dennis EschweilerPaul EhrlichSven Nebelung

Lab for Artificial Intelligence in Medicine  ·  October 2026

Code Morphometry paper Morphometry project
A recording of the running system with the demo guideline corpus. The interface is in German, and sped-up parts are marked.

Abstract

Many measurements on musculoskeletal radiographs, such as patellar height or tibial slope, are still taken by hand and then compared with reference ranges from the literature. Language models could draft such reports, but they tend to produce numbers and thresholds of their own. We present an agent that leaves both to dedicated services. A vision-language model assesses the radiograph and runs the case. A morphometry service places the anatomical landmarks and computes every measurement. A guideline search supplies every reference range, and a value is only interpreted when a retrieved passage defines its range. The model then writes a structured report in which every value can be traced to its construction on the image and every range to its source. The radiologist can ask questions, edit the impression and correct landmarks, and the report follows. Each service can be replaced and run on local infrastructure, so radiographs do not have to leave the hospital. This is ongoing work, and a reader study with radiologists is in preparation.

Method

The agent's steps in the chat. It assesses the image as a lateral view of the knee and starts the morphometry

1Image assessment

The language model looks at a downsampled copy of the radiograph and names body region and projection. The radiologist can change both before the measurements start.

Lateral knee radiograph with the landmarks and the construction of the Insall-Salvati index

2Morphometry

The morphometry service matches the image against annotated reference radiographs, places the landmarks and computes every measurement of the view. The constructions are drawn on the image.

The report with the measured values. Below the tibial slope, the guideline passage its reference range comes from

3Guideline search

For each measurement the agent searches the guideline corpus. Only a value with a retrieved reference range is interpreted, and the range links to its passage. Other values are reported as measured.

The chat. The radiologist asks about the abnormal values and the agent answers with the retrieved range

4Report and dialogue

The model drafts a structured report and answers questions in the chat. After a landmark correction the measurements update, and the agent checks the report against the new values.

Reader study In preparation

A reader study with radiologists will evaluate the drafted reports. Results will be added here.

Who supplies what

The language model plans the case and writes the text, but it never supplies a number or a threshold of its own. Every measurement comes from the morphometry service, and every reference range comes from the guideline search. A value without a retrieved range is reported but not interpreted. The radiologist can also edit the report directly, and later updates keep their wording.

The radiograph goes to the language model. The model calls the morphometry service, which places the landmarks and returns the measurements, and the guideline search, which returns the reference ranges. It then writes the report, in which the numbers come from morphometry and the ranges from the guidelines. The radiologist types a question and also edits the report directly. 13.9° Radiograph Morphometry service landmarks and measurements Guideline search reference ranges with sources Language model runs the case, writes the text Report draft Insall-Salvati index 1.1 (0.8 – 1.2) Caton-Deschamps index 0.9 (0.6 – 1.2) Posterior tibial slope 13.9° ↑ (5° – 10°) IMPRESSION The posterior tibial slope is increased. The patellar height indices are normal. No joint effusion. Radiologist asks, edits the report and corrects landmarks Why is the tibial slope flagged? writes in the impression directly knee, lateral 1.1 · 0.9 · 13.9° “posterior tibial slope” 5° – 10°

Swipe to see the whole diagram.

Numbers from the morphometry service Ranges from the guideline search Text from the language model Edits by the radiologist

Supported views

The agent reads the supported views and their measurements from the morphometry service when it starts, so a new view of the service needs no change to the agent. The reference implementation currently covers six views.

Knee lateral

Insall-Salvati ratio · modified Insall-Salvati ratio · Caton-Deschamps index · Blackburne-Peel ratio · patella morphology ratio · posterior tibial slope (three reference axes)

Knee axial

Sulcus angle · congruence angle · patellar tilt angle · lateral patellofemoral angle

Knee AP

Medial and lateral joint space width

Foot AP

Hallux valgus angle · intermetatarsal angles 1–2 and 1–5

Foot lateral

Medial arch angle · talus–first metatarsal angle · calcaneal inclination angle

Shoulder AP

Critical shoulder angle · lateral acromion angle · acromial index · acromiohumeral interval

Citation

A paper on the agent is in preparation. Until then, please cite the morphometry framework.

@article{eschweiler2026landmark,
  title   = {An Artificial Intelligence Framework for Universal Landmark Matching and Morphometry in Musculoskeletal Radiography},
  author  = {Eschweiler, Dennis and Cornejo Merodio, Eneko and Barajas Ordonez, Felix and Lichev, Aleksandar and Ignatova, Nikol and von der St{\"u}ck, Marc Sebastian and Kuhl, Christiane and Truhn, Daniel and Nebelung, Sven},
  journal = {European Radiology},
  year    = {2026},
  doi     = {10.1007/s00330-026-12555-y}
}