A Fourier domain spherical convolutional neural network for brain tissue microstructure imaging via diffusion MRI
Diffusion Magnetic Resonance Imaging (dMRI) is a non-invasive and in-vivo imaging technique tailored for tissue examination at a microscopic scale. Consequently, it is essential in the analysis of tissue microstructures of the central nervous system. To explain the measured signals, a number of biophysically inspired multi-compartment (MC) models have been proposed. They represent dMRI data as a linear combination of signals coming from different tissue compartments such as intra- and extra-axonal spaces, gray matter, cerebrospinal fluid, tumorous cells, etc. Multiple studies have shown that the parameters associated with some of these models have potential in the evaluation of several neurological diseases and in the characterization of early age brain development. However, estimation of these parameters via standard non-linear optimizers which include Levenberg-Marquardt and Gauss-Newton algorithms, often require a high number of sampling points and/or are computationally demanding, which limits their clinical application. Since in our work, we are considering dMRI signals acquired on spheres, to address the problem of microstructure parameter estimation, we propose a spherical CNN model with fully spectral domain convolutional and non-linear layers and with rotation invariant power spectrum features. In addition, the model takes into account the real nature of dMRI signals, uniform random distribution of sampling points and important noise which affects these signals. The proposed model is evaluated quantitatively and qualitatively on the problem of Neurite Orientation Dispersion and Density Imaging (NODDI) and Spherical Mean Technique (SMT) parameter estimation. The model is positively evaluated on the real data from Human Connectome Project (HCP) database and on the synthetic data generated by dmipy toolbox.
Approximating shapes in images with low-complexity polygons
We present an algorithm for extracting and vectorizing objects in images with polygons. Departing from a polygonal partition that oversegments an image into convex cells, the algorithm refines the geometry of the partition while labeling its cells by a semantic class. The result is a set of polygons, each capturing an object in the image. The quality of a configuration is measured by an energy that accounts for both the fidelity to input data and the complexity of the output polygons. To efficiently explore the configuration space, we perform splitting and merging operations in tandem on the cells of the polygonal partition. The exploration mechanism is controlled by a priority queue that sorts the operations most likely to decrease the energy. We show the potential of our algorithm on different types of scenes, from organic shapes to man-made objects through floor maps, and demonstrate its efficiency compared to existing vectorization methods.