Fig. 2
From: Coronary artery segmentation in angiographic videos utilizing spatial-temporal information
Network architecture. The input layer of the network is a 3D convolutional layer which has a kernel size of (2N+1)×3×3, and it accepts 2N+1 adjacent frames and outputs 2D feature maps of 16 channels, which are further fused by a 2D convolutional layer. The subsequent part of the network is analogous to 2D CE–Net. The network outputs a single mask image corresponding to the central frame of the input. In practice, N is usually set to 1 or 2