Zhijin Li


Stochastic texture modelling in x-ray breast imaging.

In the context of x-ray breast imaging (mammography, digital tomosynthesis, breast CT etc.), texture modelling plays an important role. Breast texture models aim to appear similar to the breast anatomical background as perceived in x-ray imaging. The requirements for such models depend on the imaging modality and on the task itself. For example, in the context of clinical discrimination tasks in mammography, a simple uniform background without any spatial correlation will be sufficient. However, other performance evaluation tasks, such as evaluation of different tomosynthesis reconstruction algorithms require backgrounds that are statistically coherent to the characteristics of clinical images. Apart from the requirement for realism, the texture models also need to be mathematically traceable. In fact, the complete characterization of a texture model is essential for mathematical observers to theoretically evaluate the performance index of a certain clinical task.
Over the last two decades, various breast texture models in 2D and 3D have been proposed. As technology and peoples understanding of breast anatomy advance further, a clear course of evolution can be observed. More and more texture models are proposed directly in 3D and their complexity is increasing: from stationary isotropic models to more complicated non-stationary an-isotropic models. Moreover, as more and more breast anatomical information is revealed nowadays, much of the research tend to focus on anatomically accurate models rather than pure mathematical models.
In this presentation we give a review of existing breast texture models. Their mathematical formulation and statistical characterization will be presented. Then we will briefly introduce our new model, which is based on 3D stochastic geometric theories.


Zhijin Li est en thèse CIFRE au CMLA avec GE Healthcare France. Il est encadré par Agnès Desolneux et Serge Muller.