A Spatio-Temporal Fully Convolutional Network for Breast Lesion Segmentation in DCE-MRI

Abstract

Breast lesion segmentation result has a huge impact on the subsequent clinical analysis, and therefore it is of great importance for image-based diagnosis. In this paper, we propose a novel end-to-end network utilizing both spatial and temporal features for fully automated breast lesion segmentation from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). Our network is based on a modified convolutional neural network and a recurrent neural network, and it is capable of unearthing rich spatio-temporal features. In our network, a multi-pathway structure and a fusion operator are introduced to acquire 3D information of different tissues, which is helpful for reducing false positive segmentation while boosting accuracy. Experimental results demonstrate that the proposed network produces a significantly more accurate result for lesion segmentation on our evaluation dataset, achieving 0.7588 dice coefficient and 0.7390 positive predictive value.