Segmentation précise des structures de l'articulation du genou sur des radiographies EOS de face

Abstract

Bone delimitation in radiographic images is a necessary step in computing clinical measurements. This delimitation can be difficult in certain regions where structures are close and may overlap, like articulations. The vast majority of tools available need the input of a clinician. To accelerate the analysis, we want to fully automate the bone delimitation process on radiographic images and still get an accurate result. One of the algorithm used to delineate bones is the search of a minimal path going through the most contrasted zones. This method can be inaccurate in certain scenarios (nearby structures, low contrast, etc.). We improved this method to get more accurate bone contours in the knee joint area. First, we search the femur’s contour with a new minimal path computed with the gradient and the pixels’ intensities. Then we detect all the contours of the tibia using an improved 3D minimal path. We compared the outputs of the minimal path used in SterEOS (EOS Imaging 3D reconstruction platform) with our approach on 147 knee joints on frontal EOS images. The SterEOS version segments the femur with an average error distance of 0.85±0.52mm (root mean square deviation±standard deviation) and the tibia with 1.69±1.13mm compared to an expert’s manually segmented contours. The proposed algorithm segmented the femur with a precision of 0.40±0.13mm and the tibia with 1.17±0.92mm (53% and 31% precision improvements respectively). To conclude, we modified the 2D and 3D minimal path algorithms and thus the precision of bone contours in radiographic images. Moreover, the 3D minimal path extracts both contours of the tibial plateau. This approach could be used in EOS software to improve the automatic 3D reconstructions and the computed clinical parameters.

Date	4 Jul 2018
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Carlos Vázquez (Supervisor) & Jacques A. de Guise (Co-supervisor)