Méthodes multi-organes rapides avec a priori de forme pour la localisation et la segmentation en imagerie médicale 3D

摘要

With the ubiquity of imaging in medical applications (diagnostic, treatment follow-up, surgery planning. . . ), image processing algorithms have become of primary importance. Algorithms help clinicians extract critical information more quickly and more reliably from increasingly large and complex acquisitions. In this context, anatomy localization and segmentation is a crucial component in modern clinical workflows. Due to particularly high requirements in terms of robustness, accuracy and speed, designing such tools remains a challengingtask.In this work, we propose a complete pipeline for the segmentation of multiple organs in medical images. The method is generic, it can be applied to varying numbers of organs, on different imaging modalities. Our approach consists of three components: (i) an automatic localization algorithm, (ii) an automatic segmentation algorithm, (iii) a framework for interactive corrections. We present these components as a coherent processing chain, although each block could easily be used independently of the others. To fulfill clinical requirements, we focus on robust and efficient solutions. Our anatomy localization method is based on a cascade of Random Regression Forests (Cuingnet et al., 2012). One key originality of our work is the use of shape priors for each organ (thanks to probabilistic atlases). Combined with the evaluation of the trained regression forests, they result in shape-consistent confidence maps for each organ instead of simple bounding boxes. Our segmentation method extends the implicit template deformation framework of Mory et al. (2012) to multiple organs. The proposed formulation builds on the versatility of the original approach and introduces new non-overlapping constraintsand contrast-invariant forces. This makes our approach a fully automatic, robust and efficient method for the coherent segmentation of multiple structures. In the case of imperfect segmentation results, it is crucial to enable clinicians to correct them easily. We show that our automatic segmentation framework can be extended with simple user-driven constraints to allow for intuitive interactive corrections. We believe that this final component is key towards the applicability of our pipeline in actual clinical routine.Each of our algorithmic components has been evaluated on large clinical databases. We illustrate their use on CT, MRI and US data and present a user study gathering the feedback of medical imaging experts. The results demonstrate the interest in our method and its potential for clinical use.