Noisy Softmax for Enhanced Subcutaneous Blood Vessel Recognition

Abstract In recent years, the identification of subcutaneous blood vessels, including veins, arteries, and capillaries, has garnered significant attention due to its potential applications in medical imaging and minimally invasive procedures. However, traditional methods often struggle with challenges such as low contrast, complex vascular structures, and variations in patient anatomy. In this paper, we introduce a novel approach to enhance the accuracy of subcutaneous blood vessel recognition by employing a dynamic noise injection strategy in the training process of convolutional neural networks (CNNs). Specifically, we address the issue of early saturation in activation functions, which can hinder the exploration capabilities of gradient descent algorithms. By injecting annealed noise into the network at each training iteration, we effectively delay saturation, ensuring more robust gradient propagation. This method promotes better exploration during optimization, thus improving the model's ability to avoid poor local minima and enhancing the identification of complex vascular patterns. Empirical results demonstrate the superiority of this approach in recognizing subcutaneous blood vessels across diverse imaging modalities, showing significant improvements in both generalization and robustness, without the need for extensive dataset-specific tuning.