Superficial Maxillofacial Venous Imaging Enhancement Using a Deep Convolutional Autoencoder with Wavelet Downsampling

Optical near-infrared imaging of superficial veins offers a non-contact means to visualize vascular distribution for injection surgical guidance. However, images of the maxillofacial superficial veins acquired under the nearinfrared spectrum often suffer from poor image quality with vascular structures indistinct. To resolve this issue, we propose a novel convolutional autoencoder network incorporating a wavelet downsampling module, designed to enhance vascular structures while simultaneously suppressing other non-vascular features. In conventional deep learning architectures, in local neighborhood downsampling operations often give rise to the loss of important spatial information that is essential for pixel-level prediction tasks. Due to this limitation, we introduce a Haar wavelet-based downsampling (HWD) module as an alternative to max-pooling operations. This approach effectively reduces feature map spatial resolution while better preserving structural information. The efficacy of our proposed enhancement framework was quantitatively assessed through standardized image quality evaluation metrics, comprising Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), and the Measure of Enhancement by Entropy (EME), and compared with traditional image enhancement algorithms. Research showed that the new model effectively enhances the maxillofacial vein structures without introducing additional noise and erroneously enhancing non-vascular high-frequency information.