Keywords: Convolutional Neural Networks, Echocardiography, Segmentation, Data Augmentation
TL;DR: Accumulating outputs from multiple models, or over augmentations of the test image on a single model, can improve poor semantic segmentation.
Track: short paper
Paper Type: methodological development
Abstract: The automatic segmentation of heart substructures in 2D echocardiography images is a goal common to both clinicians and researchers. Convolutional neural networks (CNNs) have recently shown the best average performance. However, on the rare occasions that a trained CNN fails, it can fail spectacularly. To mitigate these errors, in this work we develop and validate two easily implementable schemes for regularizing performance in 2D CNNs: model averaging and augmented inference. Model averaging involves training multiple instances of a CNN with data augmentation over a sampled training set. Augmented inference involves accumulating network output over augmentations of the test image. Using the recently released CAMUS echocardiography dataset, we show significant incremental improvement in outlier performance over the baseline model. These encouraging results must still be validated against independent clinical data.