Temporal-spatial feature learning of dynamic contrast enhanced-MR images via 3D convolutional neural networks

Abstract

Dynamic contrast-enhanced magnetic resonance imaging provide not only the information on the morphological features of the lesions, but also the changes of the lesion’s blood perfusion. In this paper, we propose a tensor-based temporal data representation (TTD) model and a multi-channel fusion 3D convolutional neural network (MCF-3D CNN) to extract the temporal and spatial features of dynamic contrast enhanced-MR images (DCE-MR images). To evaluate the performance of the proposed methods, we established a DCE-MR image dataset for non-invasively assessing the differentiation of Hepatocellular carcinoma (HCC). The TTD model achieves the accuracy of 73.96% for non-invasive assessment of HCC differentiation via MCF-3D CNN. Meanwhile, the 3D CNN with TTD achieves accuracy, sensitivity and specificity of 95.17%, 96.33%, and 94.00%, respectively, in discriminating the HCC and cirrhosis. Compared with the normal data representation method, the proposed TTD method is more conducive for 3D CNN to extract temporal-spatial features of DCE-MR images.