Tese de Doutoramento Ciências da Saúde 2024 Faculdade de Ciências Médicas, Universidade NOVA de Lisboa

Age-related Macular Degeneration (AMD) is a chronic, degenerative eye disorder, being the leading cause of irreversible visual impairment among the elderly population in the industrialized world. In AMD, there is a progressive loss of high-acuity vision attributable to degenerative and neovascular changes in the macula, culminating, in some cases, in legal blindness. The advanced stages of AMD are characterized by an atrophic and/or neovascular forms. Currently, there is neither a cure nor a means to prevent AMD. Thus, characterizing structural and functional changes in AMD holds significant importance, as it could enable the identification of meaningful non-invasive imaging characteristics or biomarkers that could be involved on disease progression. This, in turn, could have promising implications in clinical practice in terms of diagnostics, monitoring and treatment strategies. Herein, our global aim was to develop machine learning algorithms and imaging analysis tools to improve prediction and characterization of AMD progression forms using Optical Coherence Tomography (OCT) technology. Firstly, we assessed if certain OCT imaging markers were correlated with disease progression from intermediate to advanced AMD stages (complete retinal pigment epithelium and outer retinal atrophy or macular neovascularization). We designed a machine learning approach to assess the prediction capability of logistic regression (binomial and multinomial) models using OCT-extracted features in projecting disease progression. Correlation methods revealed statistically significant positive associations between several OCT variables with respect to disease progression, particularly incomplete retinal pigment epithelium and outer retinal atrophy, disruption of the ellipsoid zone, the presence of hyperreflective foci, their quantity, and their relationship with drusen (p<0.01). The binomial and multinomial logistic regression models yielded satisfactory results using the available database. Notably, the binomial model demonstrated a good discriminatory capability, and identified incomplete retinal pigment epithelium and outer retinal atrophy and disruption of the ellipsoid zone as the most relevant predictors (p<0.01). This research paved the way for future inquiries employing more extensive datasets and advanced machine learning techniques. The second task of this thesis was to evaluate the association between features of the type 1 MNV lesion and the choriocapillaris (CC) with the MNV radial growth patterns. We developed an image processing algorithm that analyzed and extracted quantitative variables in a radial manner from the MNV membrane, and its surroundings regarding CC characteristics. Potential correlations between MNV radial growth and each one of these variables were assessed. Additionally, machine learning algorithms were applied and subjected to feature selection and hyperparameters tunning, for subsequent analysis of their prediction power. The correlation assessment revealed an overall moderate/weak positive monotonic trend between MNV radial growth and each one of these flow deficits metrics: flow deficits relative density % and flow deficits average size, respectively. Tree-based machine learning models achieved satisfactory discrimination, with flow deficits relative density emerging as a significant predictor, highlighting the potential influence of impaired CC flow on MNV growth. Furthermore, certain image textural features, such as contrast and kurtosis of the gray level distribution, consistently demonstrated importance for these models’ performance. By incorporating new imaging features and increasing the size of the dataset, we may be able to increase the predictive power and robustness of the developed ML models. We also aimed to develop an innovative tool for vascular imaging - Enhanced Retinal Vasculature (ERV) - which complemented OCT angiography (OCTA) information by addressing the limitations of existing projection-resolved methods. This new imaging tool was developed using high resolution OCT technology and showed promising results in visualizing the retinal vascular network in comparison with projection-resolved algorithms, demonstrating a particular enhanced performance in terms of sensitivity and vessel continuity (p<0.001). ERV was designed to be a transversal tool and its application was also possible on conventional SD or SS-OCT. ERV validation on standard OCT technology was performed by clinicohistological correlation using a unique set of images, combining confocal and OCT/OCTA imaging volumes of the same eye. In this particular set, the previously mentioned performance improvements were also verified, observing a notable sensitivity increase of approximately 20 % when overlapping ERV images, in comparison with OCTA 3D PAR (supplied by Heidelberg software), with the corresponding histological images. With ERV, we improved the vascular imaging field, offering a valuable asset for researchers and clinicians to better visualize and characterize retinal vascular alterations, ultimately improving patient care and outcomes. This dissertation underscores the importance of technological advancements in unraveling AMD's complexities. While marking a step forward, it is crucial to acknowledge that further refinements and enhancements are essential to fully harness the potential of these models and tool. This work contributed to the ongoing efforts to comprehend the underlying mechanisms of AMD progression, and also aims to highlight the need and relevance for continued research in improving diagnostics, monitoring and treatment strategies for this multifaceted ocular disorder. As we continue to push the boundaries of knowledge and innovation, we move closer to a future where AMD is better understood and more effectively managed