Imposing boundary-aware prior into CNNs-based medical image segmentation

Abstract

While convolutional neural networks (CNNs) have become the first choice for the medical image segmentation, they still lack the critical ingredient of incorporating priors, such as smoothness and boundary shapes. The authors tackle the limitation by developing a novel prior that is boundary-aware in two ways: promoting smoothness without blurring object boundaries and punishing prediction errors according to boundary shapes. They bring the boundary-aware property into effect by weighting the prediction gradients and errors with the distance map. Their prior differs from previous approaches that either over-smooth boundaries or tend to produce rough boundaries. They evaluate their prior alongside the cross-entropy (CE) on a cardiac MRI dataset. Compared to CE alone, their prior improves the Dice score by 1.5% and Hausdorff distance by 53%. It also yielded a faster and more stable learning process.