This technology - comprising a deep neural network-based enhancement algorithm and an optical structure for vein visualization devices - falls within the interdisciplinary field of near-infrared medical imaging and artificial intelligence. On the algorithmic front, it employs an improved U-Net deep neural network. Building upon the traditional U-Net encoder-decoder architecture, this approach incorporates an attention mechanism (SE-Net) and residual connection modules to address the limitations of conventional algorithms - specifically, their incomplete extraction of minute and deep veins, and their insufficient suppression of artifacts. Through an adaptive anchor box design, the system precisely segments vascular regions from background tissues; combined with multi-scale feature fusion techniques, it integrates shallow-level textural features with deep-level semantic features. This significantly enhances the contrast between veins and surrounding tissues, enabling the detailed reconstruction of subcutaneous micro-veins with diameters of ≤0.4 mm, while simultaneously suppressing noise interference caused by skin folds, hair, and uneven lighting. With an algorithm inference latency of ≤50 ms, the system fully meets the requirements for real-time imaging. On the optical front, the system utilizes a dual-source configuration - comprising primary and auxiliary 850 nm near-infrared light sources - integrated into a coaxial beam-splitting optical structure. The primary source employs a surface-array design to ensure uniform illumination, while the auxiliary source provides annular fill lighting. This setup, combined with a narrow-bandpass filter (850 nm ± 8 nm) and a high-resolution imaging lens, minimizes interference from ambient light and diffuse reflection, thereby enabling the precise acquisition of near-infrared images. The imaging optical path is designed to be coaxial with the projection optical path, allowing the clear vein images - processed by the deep neural network - to be projected *in situ* directly onto the skin surface. The resulting structure is compact and adaptable for imaging various body parts, such as the back of the hand and the forearm. This solution effectively resolves issues regarding blurred vein visualization in complex scenarios - such as those involving obesity, edema, or darker skin tones - and significantly improves the clinical success rate of veinpuncture procedures. It is compatible with both portable and bedside-type vein visualization devices.

This invention proposes a vein enhancement algorithm utilizing a deep neural network that fuses CNN and LSTM architectures, alongside a multispectral optical imaging structure. This system addresses the issues of poor vascular continuity and inter-frame jitter distortion - common problems encountered during the dynamic imaging of traditional vein display devices. On the algorithmic front, a hybrid CNN-LSTM deep neural network is constructed. The CNN module is responsible for extracting vascular features and enhancing contrast within single near-infrared frames; it employs convolutional and pooling layers to progressively extract features such as vascular texture, width, and depth, while integrating Batch Normalization techniques to accelerate model convergence and prevent overfitting. The LSTM module handles the correlation of features across sequential frames, performing inter-frame smoothing on continuously captured vein images to suppress imaging distortions caused by minor hand tremors, thereby achieving stable vein image enhancement in dynamic scenarios. Furthermore, through an iterative adaptive threshold optimization process, the algorithm further improves the clarity of vascular edges. The optical structure employs a multispectral imaging lens assembly comprising an 850nm near-infrared imaging optical path and a 660nm visible-light auxiliary optical path; these dual paths capture images independently, utilizing differential processing to eliminate interference from stray light. The lens assembly features a dual-lens combination design paired with an adjustable focusing mechanism, allowing for flexible focal length adjustments based on the imaging distance and the specific body part being examined. This assembly works in conjunction with a high-frequency flickering supplementary light source to minimize the impact of diffuse skin reflection on image quality, thereby providing high-quality input images for the deep neural network algorithm. By balancing both dynamic imaging stability and algorithmic real-time performance, this invention is broadly applicable to clinical scenarios such as veinpuncture and intravenous infusion, enhancing both operational convenience and accuracy.

This solution features a vein viewer enhancement algorithm based on the Transformer-UNet deep neural network, coupled with a multi-band switchable optical architecture. It overcomes the limitations of traditional algorithms, which typically perform poorly in visualizing deep veins and abnormal vessels (such as sclerotic veins). On the algorithmic front, the system employs a Transformer-UNet deep neural network; by integrating the global attention mechanism of the Transformer with the local feature extraction capabilities of the UNet, it can rapidly capture the global topological structure of the vasculature while precisely extracting the faint features of deep veins. Furthermore, the introduction of an adaptive loss function - which fuses Dice loss with cross-entropy loss - addresses issues related to the low pixel proportion of vessels and the imbalance between positive and negative samples. This enhances the accuracy of vessel segmentation and enhancement, enabling effective differentiation between veins and surrounding soft tissues, as well as between vessel branches and main trunks, thereby achieving clear visualization of veins across varying depths and diameters. On the optical front, the system utilizes a multi-band switchable light source (featuring a dual-band configuration of 850nm and 940nm near-infrared wavelengths) paired with a condensing optical system. An electronic switching module facilitates rapid toggling between these two bands, allowing for the acquisition of vein images under different wavelengths to serve as inputs for the deep neural network; by leveraging the differential light absorption properties of tissues across these distinct wavelengths, the system further improves vessel discernibility. Additionally, the optical system features an anti-reflective coating design to minimize light path loss. When combined with a high-sensitivity image sensor, this design enhances imaging quality in low-light environments while ensuring adaptability to patients of different age groups (adults and children) and varying vascular conditions. This solution offers high imaging precision and strong adaptability, effectively meeting the vein visualization requirements of complex clinical scenarios while providing clear imaging support for the auxiliary diagnosis of venous diseases.