Abstract

Pulmonary hypertension (PH) is a heterogeneous multifactorial syndrome that overlaps multiple clinical classifications. It normally follows a rapidly progressive clinical course and thereby has high potential of causing cardiac failure. In the UK, it has a prevalence of 97 cases per million with standardised death rates between 4.5 and 12.3 per 100,000. Cardiac magnetic resonance (CMR) is recognised as the gold standard for imaging of the heart. In particular, it is a promising modality for automated survival prediction in PH. Accurate segmentation of CMR images is a fundamental step in predicting PH survival.

In this talk, we introduce a novel and accurate method for segmentation of PH CMR images. The method explicitly takes into account the image features learned from a deep neural network. To this end, we estimate joint probability maps over both region and edge locations in CMR images using a fully convolutional network. Due to the distinct morphology of PH hearts, these probability maps can then be incorporated in a single nested level-set optimisation framework to achieve multi-region segmentation with high efficiency. The proposed method obviates the need for level set initialisation, a common drawback associated with the local minimum nature of level set optimisation. This method is therefore fully automated. We show results on CMR cine images and demonstrate that the proposed level set method supervised by deep convolutional networks leads to substantial improvements for CMR image segmentation in PH patients.