Automatic Segmentation of 3D Large Field of View OCTA Skin Volumes using Deep Learning-based Methods

The skin comprises an intricate microvasculature vital for its function and health. With the ability to image capillary blood flow and internal structure of the skin in vivo, Optical Coherence Tomography-Angiography (OCTA) can offer new perspectives on the underlying causes and dynamic changes of skin conditions. OCTA can be used for non-invasive monitoring of disease progression and treatment, making it useful for diagnoses and treatment evaluation. To make OCTA suitable for basic and clinical research in dermatology, such as the assessment of wound healing, diagnosis and treatment of Chronic Venous Insufficiency (CVI), basal cell carcinoma, psoriasis, it is necessary to reliably analyze many images. Therefore, accurate vessel segmentation is needed. Manual segmentation of 3D OCTA images of skin is challenging, time-consuming and prone to errors, especially for small vessels that are difficult to identify. For this reason, automatic solutions have been presented, including both traditional and deep learning methods, with various architectures proposed. However, the majority of the studies on segmentation of OCTA images are related to ophthalmological applications, and only on the segmentation of the 2D maximum intensity projection of the volumes. It follows that the automatic segmentation of 3D OCTA skin volumes is an underexplored area of research. This gap in the research and the challenges of manual segmentations, highlight a need for the development of an automated segmentation approach to facilitate accurate and efficient analysis of skin microvasculature from 3D OCTA images. The goal of this thesis project is to address this gap by using deep learning methods to develop and evaluate an automatic segmentation algorithm for 3D OCTA skin volumes. The dataset contained OCTA volumes acquired on healthy skin and skin affected by CVI, ensuring a high variability in the vascular structure of the data. The high-resolution acquisition system allowed to obtain 512x512x96 pixels large volumes, comprising a field of view (FOV) of 10x10 mm in the lateral and 1.3 mm in the axial direction. The total 214 samples were split into training, validation and test set with a 50:25:25 ratio. A deep learning-based automatic segmentation algorithm was then proposed, which uses three DNN architectures, the 3D U-Net, 3D Res-Unet and 3D Dense-Unet, and combine them with Ensemble Learning methods to obtain connected vascular networks in the segmentations. Different loss functions (Dice, Jaccard and a connectivity loss) were also evaluated, with data augmentation and a vesselness filter on the training set, to obtain the best model for the final testing. Finally, two post-processing techniques were applied to the predictions to improve the connectivity of the vascular networks, namely the removal of small, isolated objects, and a closing operation which fill gaps and holes. The algorithm’s performances were evaluated against the segmentations obtained using the Amira semi-automatic software. The metrics obtained reach Dice Scores higher than 80% for 60% of the samples, with the presence of few outliers with low metric values, more often due to an inaccurate reference segmentation, rather than a low-quality automatic segmentation. The results are promising considering the FOV of the volumes used and the variability in the data. Future works might include the automatic classification of each volume based on the computation of vascular metrics, to further validate the proposed automatic segmentation methods.