Title:	Generative AI and foundation models for atypical femur fractures
DNr:	Berzelius-2025-59
Project Type:	LiU Berzelius
Principal Investigator:	Anders Eklund <anders.eklund@liu.se>
Affiliation:	Linköpings universitet
Duration:	2025-02-19 – 2025-09-01
Classification:	30208
Homepage:	https://liu.se/en/research/prio
Keywords:

Abstract

Orthopedic trauma to the extremities is the most-common reason for visits to emergency departments worldwide, and together with other musculoskeletal conditions such as back pain and osteoarthritis, fractures are the major cause for years lived with disability globally. While most musculoskeletal conditions can be treated as non-urgent, acute fractures require urgent attention to be diagnosed correctly, followed by acute treatment to decrease the risks for short- and long-term morbidity and mortality. Atypical femur fractures (AFF) represent a very rare disease pattern among the already rare stress fractures of the femur. Much attention has been given to these fractures because of the paradoxical causal relationship between the most commonly used drugs to treat osteoporosis (bisphosphonates) and a dramatically increased risk for AFF with long-term bisphosphonate treatment. With a yearly incidence of 1.1–2.2 AFF per 100,000 inhabitants in Sweden, AFF comprise only about 0.25% of all fractures of the femur. Since AFF are so rare, clinicians might overlook this fracture pattern among the abundance of other fractures. In this project we are therefore using deep learning for automatic detection of AFF from radiographs. Using our huge dataset consisting of several million images, we will train foundation models (convolutional neural networks and vision transformers) to improve the detection of AFF, compared to training with a few thousand images. To facilitate data sharing, we will also train generative models (GANs and diffusion models) to synthesize realistic synthetic images.