Level-set segmentation of 2D and 3D ultrasound data using local gamma distribution fitting energy

Ultrasound (US) data suffer from speckle noise as well as intensity inhomogeneities due to underlying changes in acoustic properties of tissue structure and/or the effects of acoustic focusing and attenuation. This paper describes a 2D and 3D variational level-set method for segmenting such data. To deal with the local statistics of speckle noise, the data term of the level-set energy function is based on local gamma distributions which have shown an ability to model envelope data and gray-level pixel intensities of B-mode clinical images. Local statistics are estimated at a controllable scale using a smooth function with a compact support, a mollifyer, and the method of moments. Compared to manual segmentation, the investigated method provides a high Dice similarity coefficient (DSC) on 3D simulated data, an average DSC of 0.915 on 12 B-mode images of murine tumors acquired with a clinical US system, and average DSCs of 0.920, 0.806 and 0.975 for three media of 54 3D envelope data sets acquired with a high-frequency, focused transducer from dissected human lymph nodes. It also outperforms methods that employ local Gaussian statistics instead of local gamma statistics.