Postdoctoral position: Deep Learning for cardiac segmentation
12 Avril 2022
Catégorie : Post-doctorant
A 12-months postdoctoral position is proposed on cine-MRI 3D segmentation of the left ventricle using Deep Learning guided by geometric parametric models. This project will be supervised by Laurent Sarry (Institut Pascal, Clermont-Ferrand) and Patrick Clarysse (CREATIS, Lyon).
Position description: drive.uca.fr/f/b64c0e7a43484ef89cc9/
Lab: Institut Pascal UMR 6602 - UCA/CNRS
Head: Evelyne Gil
Supervisors: Laurent Sarry (Institut Pascal, PU UCA), Patrick Clarysse (CREATIS, DR CNRS)
Location: axe TGI, équipe CaVITI, Faculté de Médecine, BP 38, 63001, Clermont-Ferrand
The post-doc project concerns the development of segmentation tools for cine MRI images to facilitate the estimation of 3D myocardial function. Cine MRI has become the reference modality in terms of accuracy and reproducibility for the estimation of myocardial anatomical features (volume, mass, thickening...), as well as for global function (ventricular ejection fraction). In the context of the ANR 3DStrain project (2011-2015), we have proposed a new variational formalism for the estimation of the regional function (strain), consisting in computing a dense field of deformation at each point of the myocardium. However, this estimation remains dependent on the quality of the segmentation of the endocardial and epicardial walls, which we then managed using deformable B-spline models.
The semantic segmentation methods by Deep Learning have shown very good performances, both in 2D and 3D on this type of data. We have recently confirmed this by training 2D and 3D U-Net Deep Neural Network models on our own 3T cine MRI database of which about 200 endo and epicardial volumes of the left ventricle (LV) have been manually delineated (corresponding to more than 2000 slices).
However, the standard models do not ensure an anatomically plausible segmentation, because they do not incorporate any geometric or topological shape constraints. For example, the following figure illustrates the difficulty of accurately segmenting MRI slices at the basal level:
Basal cine MRI slice (left), endocardium ground truth (middle), segmentation predicted with 3-D U-Net (right).
The main objective of the project is therefore to integrate a constraint in the form of geometric models into the Deep Learning framework. These models will have to make a compromise in terms of degrees of freedom to represent the anatomical variability of the myocardial walls, while sufficiently constraining the final prediction.
Another advantage of these models will be the ability to interactively correct the predicted segmentation. Indeed, the a posteriori correction of a segmentation performed at the pixel/voxel level remains very time consuming, in particular in dynamic imaging. If the prediction directly produces a parametric model that can be modified interactively, this facilitates the correction of the result.
The methodology consists in replacing the encoder/decoder structure of the semantic segmentation networks by an encoder/FCN structure allowing the regression of the parameters of the chosen geometric model. The geometric model, representing the endocardial or epicardial surface of the LV, must be locally deformable, which leads to B-spline contours in 2D and to tensor product B-spline surfaces in 3D. Rather than using the control points as parameters, we will study the possibility of replacing them by the nodal points which have the advantage of belonging to the contour or the surface and thus being more easily editable a posteriori. B-splines also have the advantage of being able to compute sampled points from parameters in a tensor formalism, compatible with frameworks such as TensorFlow or PyTorch in order to benefit from GPU or TPU hardware accelerations.
Because of the particular geometry of the LV, it is possible to simplify the model with a single B-spline of the radial distance to the major axis of the left ventricle, rather than considering one B-spline per coordinate. This simplified representation assumes the ability to normalize the orientation of the acquisitions around the major axis.
The activities of the recruited post-doc will follow an incremental approach of increasing complexity, with deliverables resulting from network learning and predicted segmentation on the same test basis at each step: