1Université de Lyon, CREATIS - INSERM, Lyon, France
The objective of this paper is to present an overview of the current segmentation techniques applied to ultrasound medical images. In the associated talk, we propose to answer some basic questions: What are the specific difficulties raised by ultrasound imaging? What are the main methodological approaches? What is a "good" segmentation? What perspectives can be drawn from the evolution of ultrasound imaging techniques?
The specificities of ultrasound image formation make them very challenging to segment: diffusion yields speckle noise, attenuation and diffraction yield contrast variation, image properties are orientation dependent and the field of view is often limited. Moreover the ability of ultrasound imaging to perform real-time acquisition sets challenges on the implementation of the algorithms.
A wealth of techniques has been proposed to address segmentation of ultrasound images. Whatever the underlying formulation, these approaches broadly share two common aspects. First, they rely on the proper choice and extraction of image ("low-level") features/properties that characterizes the object to be detected (gradient, texture, etc.). These features alone can however hardly provide reliable segmentation. The second aspect thus consists in introducing a priori information about the object to be detected, i.e. constraints such as shape, volume, motion. Assessing the quality of a segmentation algorithm is particularly difficult in medical imaging, because defining a ground truth is not straightforward. This evaluation may resort to physical phantom imaging or numerical simulation. The currently available numerical simulations are however not mature enough to serve as a basis for segmentation, and are mainly used as a first prototyping step. Validation is thus usually still performed by comparing the segmentation provided by the algorithm to the segmentation performed manually by medical experts. We will present in this talk various example borrowed from the field of cardiac imaging, vascular imaging and prostate and breast cancer imaging to illustrate the principle, results and shortcomings of cureent segmentation techniques.
One of the most interesting perspectives consists in exploiting new image-based features. To date, the vast majority of the approaches are based on log-envelope images. We will discuss the possibility of processing raw RF data, which has been almost unexploited till now. Another perspective concerns the evaluation of segmentation. The availability of intensive computing resources such as distributed computing infrastructures can provide a way toward more realistic simulations as well a basis for algorithm prototyping. We will exemplify the use of such a validation platform and show some preliminary results.