AI turns routine CT scans into osteoporosis screening

Osteoporosis often develops unnoticed for years and may only be diagnosed after a fracture or the onset of back pain. Researchers in Germany are developing AI-based software that could help identify reduced bone density earlier by extracting additional information from CT scans that patients have already undergone for other medical reasons.

The software is being developed within the BMDNow research project by the Fraunhofer Institute for Computer Graphics Research IGD, Goethe University Frankfurt and IT company garritz online media international. The project is funded by Hessen Agentur.

Bone density as an additional CT finding

Dual-energy X-ray absorptiometry (DXA) is considered the standard method for diagnosing osteoporosis. However, it is generally performed when reduced bone density is already suspected. BMDNow takes a different approach. The software analyses existing routine CT images, such as scans originally performed to examine the heart or lungs. This means that bone mineral density (BMD) could potentially be assessed without an additional imaging examination.

“We run our algorithm on the available images and obtain a bone density analysis essentially as a byproduct,” explains Stefan Wesarg, a research scientist at Fraunhofer IGD. According to the researchers, this approach could avoid additional radiation exposure and reduce the time required from healthcare professionals. The system combines neural networks for automated image segmentation with imaging techniques designed to calculate the bone density of individual vertebrae.

AI analyses vertebral tissue

The software first identifies the trabecular region inside the vertebral bodies, the spongy tissue within the bone. The researchers tested several AI models under real-world conditions using patient data from Frankfurt University Hospital. The algorithm then creates a three-dimensional representation of bone density distribution within this region. It analyses components including fat and calcium and visualises their proportions using colour coding. Bone density is calculated using calibration curves previously created with reference phantom models.

The resulting measurements are intended to indicate whether a patient has osteoporosis or osteopenia, a condition characterised by lower-than-normal bone density that can precede osteoporosis. Earlier detection could allow physicians to consider treatment before fractures occur.

The researchers currently use dual-energy CT, which acquires images at two different X-ray energy levels. Their next goal is to adapt the technology for conventional single-energy CT scans, potentially increasing the number of existing scans that could be analysed. The project partners are also working towards a reference database and are continuing to train, optimise and validate the software prototype. Medical device approval is targeted within one to two years, meaning the technology is still under development and not yet available for routine clinical use. BMDNow took second place in the 2025 Hessen Ideen competition. The project was recognised for its potential to make better use of existing medical imaging while reducing the need for additional examinations.

3D AI camera

Earlier this year, researchers at the Fraunhofer Institute for Manufacturing Engineering and Automation IPA, announced that they are developing an AI-based system to automate the documentation of ultrasound examinations. The SonoMap project combines a 3D camera with AI image processing to determine where an ultrasound image was taken and the angle of the probe relative to the patient’s body. Currently, doctors often have to document this information manually, a process that can take considerable time and may introduce inaccuracies. These differences can complicate follow-up examinations, particularly when monitoring abnormalities such as tumours or cysts.

SonoMap uses a 3D camera to detect the ultrasound probe while mapping the patient’s body surface. AI algorithms determine the probe’s position and orientation and link these data to a simplified 3D body model. The resulting interactive visualisation can be stored and viewed from different perspectives, potentially making follow-up examinations more consistent while reducing documentation time for clinicians.