Non-stationary deep lifting with application to acute brain infarct segmentation
We present a deep learning based method for segmenting acute brain infarcts in MRI images using a novel input enhancement strategy combined with a suitable non-stationary loss. The hybrid framework allows incorporating knowledge of clinicians to mimic the diagnostic patterns of experts. More specifically, our strategy consists of an interaction of non-local input transforms that highlight features which are additionally penalized by the non-stationary loss. For brain infarct segmentation, expert knowledge refers to the quasi-symmetry property of healthy brains, whereas in other applications one may include different anatomical priors. In addition, we use a network architecture merging information from the two complementary MRI maps DWI and ADC. We perform experiments on a dataset consisting of DWI and ADC images from 100 patients to demonstrate the applicability of proposed method.