Near-infrared vascular image segmentation using multi-stage enhancement and TAU-Net

Abstract

Near-infrared optical imaging technology provides a non-invasive solution for visualization and monitoring of subcutaneous vascular structures. In order to solve the problems of low vascular image quality and inefficient and inaccurate manual segmentation, we propose a complete set of image processing methods. First, the blood vessel images are preprocessed by the background removal, Gaussian filtering, and contrast stretching. Then the image details are enhanced by a multi-stage enhancement method, which combines the Residual Convolutional AutoEncoder (RCAE) and Contrast Limited Adaptive Histogram Equalization (CLAHE) to effectively improve the contrast between vascular region and other tissue regions. Finally, the images are segmented by our Triplet Attention U-Net (TAU-Net) model, which improves the efficiency and performance of attention mechanism. The TAU-Net introduces a triple attention module in U-Net for the first time, which strengthens the computational ability of spatial and channel attention models. The main segmentation head and auxiliary segmentation head are combined to improve the gradient information, promote the multi-scale learning of network. Numerous experimental results show that our model can flexibly process blood vessel images of various quality levels and distribution forms, and effectively segment their contours well.