3D-to-2D mapping for user interactive segmentation of human leg muscles from MRI data

The automated computation of muscle volume from MRI of human legs is an open problem in the biomedical imaging community. Such automation has the potential to provide an objective measure of effectiveness of pre- and post-surgery treatments. In this paper, we take a step toward automation by proposing a framework for user interactive segmentation of MRI of human leg muscles. Our framework is built upon the strategy of bootstrapping: after the first few tedious segmentation results are achieved, and once a small database of segmented muscles is built, user interaction is reduced. Further, as the database of segmented muscles grows, the user interaction becomes more efficient. At the heart of this proposed framework is a simple, computationally attractive 3D representation of muscles. By a generalized cylinder model, we represent a 3D human leg muscle by two smooth 2D images, which enables application of 2D image processing and analysis methods in this complex multi-segment 3D problem. The smoothness of a leg muscle is modeled by the smoothness of the 2D images. Interdependence and relative positions of leg muscles are represented by a linear combination (basically, convolutions) of such 2D images. We demonstrate that fitting and editing of these models during user interactive segmentation of MRI data are computationally efficient, because our linear interaction model can be represented in the Fourier domain.