A Hybrid Deep Learning Architecture for Vessel-Enhanced Denoising in OCTA Images

High-quality retinal images are essential for accurately diagnosing and treating ophthalmic diseases. Optical Coherence Tomography Angiography (OCTA) images, reconstructed from multiple OCT B-scans, provide detailed visualizations of blood vessel distribution. However, their quality is often degraded by inherent speckle noise introduced during vascular signal extraction, making it challenging to assess normal and abnormal retinal blood vessels. This paper presents VGG16-TUNet, a novel hybrid deep-learning architecture designed for vessel-enhanced OCTA denoising. It enhances the standard UNet by integrating three key components: a pre-trained VGG16 encoder for hierarchical feature extraction, a transformer bottleneck for capturing low-scale features and long-range dependencies, and Squeeze-and-Excitation(SE) blocks at multiple decoder levels for channel-wise feature recalibration. We validate our model on the UT-FSOCTA dataset across macular and optic disc-centered OCTA images. VGG16-TUNet achieves PSNR values of 24.285 dB (macular) and 24.689 dB (optic disc), demonstrating its effectiveness in improving OCTA image quality and clinical assessment.