Multitasks convolutional networks for medical image segmentation and regression.

The analysis of medical images consists of a number of tasks, some of which have significant mutual interaction. For example, predicting relative in-vivo pressures from velocity field measurements (we refer to this tasks as pressure prediction or simply regression) requires a preliminary characterization of the fluid domain through segmentation. In turn, image segmentation quality can be improved based on flow features, such as spatial velocity gradients and acceleration that can be captured by trainable convolution kernels. Therefore, instead of treating these two tasks independently, we investigate the accuracy of a number of multitask architectures designed to promote the exchange of information between the segmentation and regression tasks. We train these networks using realistic velocity fields, with a random field noise model resulting from the non linear reconstruction of undersampled single-coil k-space acquisitions. Finally, we consider various forms of physics-based regularization designed for convolutional network architectures.