A deep learning approach for the analysis of voxel composition in magnetic resonance imaging

This work proposes a machine learning algorithm for analyzing the composition of each voxel in Magnetic Resonance Imaging framework. Since the voxel defines the spatial resolution limit of the system, conventional imaging approaches are not able to distinguish an homogeneous voxel (composed by a single tissue) from an heterogeneous one (composed by several tissues). It causes loss of useful diagnostic information not allowing to detect tissues smaller than the voxel dimension. This paper proposes a neural network detection algorithm that, exploiting several acquisitions of the same voxel, is able to distinguish heterogeneous voxels from the homogeneous cases. A synthetic dataset was designed for the neural network training: homogeneous and heterogeneous voxels were simulated and several imaging sequences have been considered to acquire voxel signals. Results seem to be promising, in particular regarding on its application in a real clinical context: in order to analyse the voxel, the algorithm requires a reduced number of acquisitions (that means reduced time for the clinical examination) and its low computational complexity allows a quick analysis of all voxels of the examined body slice, as proved by the algorithm validation through a test on a real magnetic resonance data.