Subcutaneous Blood Vessel Boundary Detection Using Deep Convolutional Networks

Abstract Edge detection in subcutaneous vascular imaging is an essential task for various medical applications, particularly in the context of non-invasive diagnostics. While convolutional neural networks (CNNs) have made significant strides in addressing edge detection problems, existing approaches may struggle with the unique challenges posed by the intricate, high-variability structures of vascular networks, especially when considering variations in scale, shape, and intensity. In this paper, we introduce a novel edge detection technique tailored to subcutaneous vascular imaging, leveraging richer convolutional features (RCF) to enhance performance. RCF aggregates features from multiple layers of the network, creating a more discriminative representation that effectively captures the multiscale and hierarchical nature of vascular structures. This method enables a comprehensive, holistic image-to-image prediction, which is particularly beneficial for delineating complex vascular boundaries. The proposed technique utilizes a standard CNN architecture, such as VGG16, and is optimized for robust backpropagation-based training. We show that this approach can accurately detect vascular boundaries, outperforming traditional methods in both precision and robustness while maintaining efficient processing speeds. Furthermore, the versatility of RCF is demonstrated through its potential application to other related medical imaging tasks, such as segmentation of vascular regions in images.