Förklaringsbar Automatisk Klassificering av Kotkompressioner från Rutinmässiga CT-undersökningar : Möjliggör Tidig Diagnostik av Benskörhet Genom Geometribaserade Referenspunkter i Rutinmässig CT-diagnostik

Hip fractures are among the most common fracture types in Sweden and can have severe consequences for a patient. Early indications of increased risk could help in prevention.Vertebral compression fractures are a strong indicator of increased hip fracture risk. Compression fractures could be used as early risk signs, but they are underdiagnosed. Automatic classification on CT images could aid radiologists and clinicians in noticing and reporting these fractures.Therefore, the objective of this thesis is to develop an automatic, explainable solution for vertebral compression detection andgrading on CT images. An automatic geometry-based landmark placement solution was created according to the clinical standard GENANT classification system. Two rule-based models derived from measured distances from paired landmarks were developed, and a machine learning model was trained. Explainable AI tools were applied to the machine learning model. The work was performed using the accessible VerSe dataset of 1,180 non-fractures and 289 fractured vertebrae. The machine learning model showed an overall accuracy of 80%, which was similar to previously presented solutions. Unlike most other models, the developed tool also classified the type and degree of fracture, closely following the GENANT classification system. The results from the explainable AI tools suggest that height deviations from the average were a common differentiation for fractures. The automatic solution performed well in determining vertebral compression fractures on CT images, and it could potentially be used to predict future hip fractures from opportunistically acquired CT images.