Investigation of the added value of 3D segmented CT images to the classification accuracy of proximal humeral fractures (FractSeg)

ARI Exploratory Research

J Dauwe, K Mys, J Schader, B Gueorguiev, P Varga


Osteosynthesis of proximal humeral fractures remains challenging with a high failure rate reported in the literature. This might be due to the complexity of these injuries, the difficulties in the appropriate selection and correct execution of treatment. Understanding and correct classification of the fracture are important for preoperative planning. Nevertheless, this is a challenging task in case of complex fractures, partially due to difficulties in recognizing their 3D extent, including the number and dislocation of fragments. Preoperative CT is a standard procedure in most hospitals, but the contained information may not be fully utilized. Advanced visualization of the CT images may improve the accuracy of fracture classification.




To investigate the feasibility and added value of semi-automatic segmented 3D CT visualizations in proximal humeral fracture classification for observers with two different experience levels: residents and specialized shoulder surgeons.



Seventeen patients with proximal humeral fractures and a preoperative CT scan were included in this retrospective study. The CT scans were semi-automatically segmented, indicating every fracture fragment in a different color. Fracture classification ability of 21 orthopedic residents and 12 experienced shoulder surgeons was tested. Both groups were asked to classify the fractures using 3 different modalities (standard slice-wise CT analysis, conventional 3D CT reconstruction, and 3D segmented model) into three different classification systems (Neer, AO and LEGO). All participants were able to classify the fractures significantly better using the 3D segmentations (94% correct answers on average) compared with the conventional 3D reconstructions (54%) and with the standard slice-wise CT analysis (35%) into the three classification systems, p<0.01. Both observer groups achieved significantly worse classification accuracy in the LEGO system compared to the two others.

  • Partner

    Putzeys G (MD), AZ Groeninge Hospital Kortrijk, Belgium

    Nijs S (Prof), University Hospitals Leuven, Belgium


Cookies help us improve your website experience.
By using our website, you agree to our use of cookies.