Multigrid level-set segmentation of high-frequency 3D ultrasound images using the Hellinger distance

We propose a multigrid level-set segmentation algorithm for the segmentation of 3D high-frequency ultrasound images. Target applications include the quantitative analysis of lesions or normal structures within cutaneous and superficial subcutaneous tissues. The method is based on an a non-parametric region-based cost function, the Hellinger distance, which is related to the Bhattacharyya coefficient. The choice of this as a cost function allows the discrimination of different tissues using the statistics of the signal. Unlike other methods, it is also applicable when tissues are heterogenous. Moreover, the choice of a nonparametric method lends itself to a multigrid approach, which allows significant gains of speed, a critical property for 3D images. We show examples of segmentation of tumors and dermis in both realistic simulated images and clinical images from the Dermcup 25MHz skin probe (Atys Medical).