Fast 3D segmentation

Problem and context

Fast segmentation of 3D organs is a strongly needed step in many clinical applications. Our goal is to developed an efficient framework well adapted for the accurate segmentation of anatomical structures in medical imaging, with a particular interest for the left ventricle of the heartBARB-13BARB-13bBARB-14QUEI-14ALME-16.

Methods, contributions and results

We have introduced an novel variational segmentation framework which allows performing near real time segmentation of 3D anatomical structures. This method, named B-spline Explicit Active Surface (BEAS), corresponds to a variational active surface formalism strongly inspired by the classical level-set method BARB-12a. The main idea resides in exploiting the equivalence, under certain conditions, between implicit and explicit surface representation. The main interest of such approach resides in its intrinsic capacity of resolving segmentation problems into a space with lower dimensions, making the algorithm highly ompetitive in terms of computation costs. Moreover, in order to make the algorithm even more faster, we expressed the underlying explicit surface through a B-Spline representation. Thanks to the corresponding formulation, we showed that the evolution of the deformable surface corresponds to a simple succession of separable convolution products, which ensures even more the high performance of the proposed algorithm in terms of computation time. Figure 1 shows an illustration of the proposed framework.

Figure 1 Three iteration steps of the algorithm segmenting a 3D squirrel. (t0) initialization, (t1) intermediate step, (t2) final result at convergence.

The evaluation of our method as part of the Challenge on Endocardial Three-dimensional Ultrasound Segmentation (CETUS challenge, MICCAI 2014) has shown that it ranks first in the automatic segmentation category based on the analysis of 30 patients over the 5 automatic methods which have been evaluated BERN-15a. Figure 2 shows an example of a segmentation result obtained using the proposed BEAS framework on a dataset provided during the CETUS challenge.

Figure 2 Top row: segmentation of the end-diastolic phase (Mean absolute distance error = 1.04mm). Bottom row: segmentation of the end-systolic phase (Mean absolute distance error = 1.69mm).