Towards Robust Cardiac Segmentation using Graph Convolutional Networks - real time demo

Fully automatic cardiac segmentation can bea fast and reproducible method to extract clinical measure-ments from an echocardiography examination. The U-Netarchitecture is the current state-of-the-art deep learningarchitecture for medical segmentation and can segmentcardiac structures in real-time with average errors compa-rable to inter-observer variability. However, this architecturestill generates large outliers that are often anatomicallyincorrect. This work uses the concept of graph convolu-tional neural networks that predict the contour points ofthe structures of interest instead of labeling each pixel.We propose a graph architecture that uses two convolu-tional rings based on cardiac anatomy and show that thiseliminates anatomical incorrect multi-structure segmenta-tions on the publicly available CAMUS dataset. Additionally,this work contributes with an ablation study on the graphconvolutional architecture and an evaluation of clinicalmeasurements on the clinical HUNT4 dataset. Finally, wepropose to use the inter-model agreement of the U-Net andthe graph network as a predictor of both the input andsegmentation quality. We show this predictor can detectout-of-distribution and unsuitable input images in real-time.Source code is available online: https://github.com/gillesvntnu/GCN_multistructure