Enhancing Optical Coherence Tomography Angiography through GAN-based techniques

Optical Coherence Tomography Angiography (OCTA) is a powerful imaging technique that provides non-invasive visualization of vascular networks, primarily in the eye but also in the skin. By capturing multiple optical coherence tomography (OCT) B-scans at the same location over time, OCTA identifies motion contrast arising from flowing blood cells. In this thesis, we propose a novel approach to accelerate OCTA data acquisition by leveraging Generative Adversarial Networks (GANs). The primary objective is to generate high-quality OCTA enface images from lower-quality inputs, which are typically derived from only two OCT volumes, compared to the higher-quality images used as ground truth that require four OCT volumes. The dataset utilized in this thesis includes images obtained from skin samples in both healthy individuals and patients diagnosed with Chronic Venous Insufficiency (CVI). The chosen format for the study is the Median Intensity Projection (MIP) of 5 slices, obtained in the enface plane. To ensure robustness and generalizability, the dataset is properly partitioned into training (80%), validation (10%), and test (10%) sets. This work presents two different approaches to image quality improvement of OCTA using GAN: super-resolution GAN (SRGAN) and pixel2pixel GAN. The SRGAN proposed is more specifically an Enhanced Super-Resolution GAN wherein the input and output images are of the same resolution allowing enhancement of image quality. This methodology integrates a Residual-in-Residual DenseNet as a generator with a Patch GAN serving as the discriminator. In addition, the introduction of the VGG19 model further enhances perceptual quality by aligning generated images with high-level features extracted from the reference image. The Pixel2Pixel GAN (Pix2Pix) architecture employing a U-Net as its generator represents a robust framework for image translation from low to high-quality versions. The U-Net structure facilitates precise reconstruction by capturing fine details from the low-resolution input and transforming them into high-resolution outputs with enhanced fidelity. Complementing this, a PatchGAN discriminator assesses local image patches to ensure the generated images exhibit realistic textures and structures akin to authentic high-quality counterparts. In addition to visual assessment, objective metrics such as PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index), and MSSSIM (Multi-Scale Structural Similarity Index) were utilized to quantitatively evaluate image quality. Across all evaluations, MSSSIM exceeded 90%, indicative of high fidelity in image reconstruction. PSNR values hovered just below 30 dB, further affirming the overall quality of the reconstructed images. In conclusion, our proposed methods utilizing SRGAN and Pixel2Pixel GAN manage to speed up acquisition times as it manages to achieve good image quality even with the acquisition of a lower number of OCT volumes.