Détection semi-automatique des contours des côtes à partir de radiographies thoraciques postéro-antérieures de patients scoliotiques

The method for stereoradiographic 3D reconstruction of the rib cage currently used at CHU Sainte-Justine research center implements a completely manual 2D detection step done by a technician. This manual detection method, however, has weaknesses which limit its clinical applications. The current research project's main objective is to develop and implement a novel semi-automatic method for 2D detection of actual scoliotic rib borders that would be accurate and robust whereas being applicable to both dorsal and ventral portions of all 24 ribs of the rib cage, using both postero-anterior views PA-0° and PA-20°.

Literature on that matter contains a lot of fully automatic methods for detecting actual rib contours, but they limit themselves to detecting normal, non-scoliotic ribs in standard zero-angle thoracic radiographs. In addition, most of those methods, in order to constrain their search space, rely on shape priors stating that the ribs are oriented horizontally in the lung area or that the intercostal space is constant from one rib to another or even that the rib cage morphology can be predicted by a morphological statistical model with a limited number of parameters. In the scoliosis context, however, all of these priors become invalid or inappropriate. This is the reason why a new 2D rib detection method, that is semi-automatically driven, was developed in the current project. In fact, great variations in ribcage morphologies from one scoliotic patient to another indicate that a semi-automatic detection approach with manual correction capabilities in a corrective post-processing step would be more suitable than a fully automated approach.

The proposed method is based on the fact that a radiographic rib, no matter how bent it is, will always present relatively parallel upper and lower borders. The method uses an edge-following approach with multiple-path branching. For each rib to detect, the user has to insert four starting points, which are to be located at both rib extremities, on both upper and lower borders. Then, four edge-following processes, initiated from these four starting points, each form an edge-following tree graph by the end of the algorithm. The novelty of the method is its ability to follow multiple promising paths simultaneously. The final rib border solution is obtained by processing the four edge-following trees and by retaining the most parallel pair of borders (upper and lower). Besides, in the semi-automatic context, we also designed manual correction of the ribs as a post-processing step. The user is then able to manually delineate ribs that were missed during edge-following execution, or can manually correct those which seem to have been incorrectly detected.

The method was tested on 44 chest radiographs of scoliotic patients for which all the ribs were a priori marked by an expert. Also, each of these radiographs were classified beforehand as good, regular or poor quality, by taking many image characteristics into account. Analysing the results, we note that the method is well suited for actual border detection of scoliotic ribs. The average 2D detection accuracy is 2,64 pixels. When compared to previous studies in literature, the proposed method is more accurate and flexible. Also, what comes out as a great asset is its ability to detect both ventral and dorsal parts of the ribs, whereas being applicable to all the 24 ribs of the rib cage. Besides, when considering the different quality levels separately, the accuracy changes from 3,60 pixels, to 2,71 pixels and to 1,97 pixels for poor, regular and good quality levels, respectively. Analysing the robustness, we observe that 93% of all the marked ribs were effectively detected, among which 27% required minor manual corrections in a post-processing step and 5% required major corrections by the user. 61% of the detected ribs did not require any correction and their associated average 2D accuracy is 1,5 pixels. The average user time for marking all the ribs on a pair of radiographs is less than 12 minutes and the one for manual correction step, about 17 minutes. The proposed method is thus 10 times faster than the previous manual one, or 4 times faster when considering the manual correction step.

The main contribution of this work is to give us the ability to accurately detect the actual borders of scoliotic ribs instead of limiting ourselves to their midlines. By enhancing the 2D detection accuracy and repeatability inside the 3D reconstruction work frame, the present study directly contributes to the efforts done towards decreasing radiation doses inflicted to patients during follow-ups, and at the same time, it calls for a totally new 3D reconstruction technique that would provide us with more personalized 3D models, considering that full information from rib borders would be used instead of just fitting generic rib 3D models onto reconstructed rib midlines. (Abstract shortened by UMI.)