Retinal Vessel Segmentation from Optical Coherence Tomography Angiography images through Deep Autoencoder architectures and Adversarial approaches

The retinal vessel segmentation is of great importance in the diagnosis of numerous diseases including diabetic retinopathy, atherosclerosis and hypertension. The structure of the vasculature is characterized by vessels of different thicknesses and lengths, and, given the strong imbalance between the vascular part and the background, this segmentation task is rather complex, especially if the specialist is not supported by Artificial Intelligence algorithms. In this work, the dataset OCTA-500 has been used to perform a supervised segmentation. The portion of the dataset that has been selected contains 2D images from the projection map of volumetric images obtained by Optical Coherence Tomography (OCT) and Optical Coherence Tomography Angiography (OCTA), with associated segmentation masks. These two imaging technologies are rapid and non-invasive techniques that allow for microscopic imaging of both structural and vascular features of the retina. In particular, OCTA can provide a qualitative and quantitative assessment of the retinal and choroidal microvascularization. After an overview of the current segmentation techniques, it has been decided to employ algorithms based on the use of deep neural networks. Following the application of several preprocessing steps, the analysis has been initially performed through the use of a standard Autoencoder architecture. Subsequently, what is considered as one of the state-of-the-art of segmentation has been applied, i.e. the U-Net structure. Moreover, a comparative analysis has been carried out with some new versions of it, such as the Attention U-Net and the Attention Residual U-Net. Finally, the potential of adversarial training has been explored to understand the characteristics of GAN architectures in the segmentation task, in particular with the use of the Pix2Pix GAN. Ad hoc metrics have been chosen in order to measure the performance of such a complex task and to train the model in the most correct way. The obtained results show a great enhancement of the performance when using the OCTA projection maps instead of the OCT ones. Indeed, for each implemented model, the Dice score related to the vascular class is characterized by an improvement of at least 6%. The best configuration found is the one related to the use of OCTA images through the U-Net network, reaching values of 85.6% for the Dice score and 86.5% for the mean IoU score.