Semantic Neural Network for Micro-Vessels Extraction

Convolutional neural network requires extensive pixel-wise annotated samples, which is imperative in the medical domain. Besides, reducing the sample size for model training alters the data distribution and causes a domain shift. The continuous pixel loss in the encoding phase and the inability to retrieve intensities in a decoding phase restrain conventional deep learning methods from extracting essential information. The article aims to resolve these issues and improve segmentation fidelity by extending the generative adversarial network. We have developed an end-to-end semantic generative adversarial network (GAN) with residual connections and dense convolutions. The residual connections helped in network depth management, preserved the useful representations, and alleviated the vanishing gradient problem. Similarly, the dense convolutions enhanced the flow of shallow features, made the training process sufficiently guide the feature selection, and avoided redundant data sharing across the network. Our method has shown promising results for DRIVE, STARE, and CHASE-DB1 datasets with an improved response in extracting micro-vessels and restoring intensities at bifurcation points.