Development of tools within a Graphical User Interface for the automatic longitudinal analysis of Dry Age-related Macular Degeneration

Tese de mestrado integrado, Engenharia Biomédica e Biofísica (Sinais e Imagens Médicas) Universidade de Lisboa, Faculdade de Ciências, 2022Age-related Macular Degeneration (AMD), a progressive neuroretinal degenerative disease, is the leading cause of irreversible vision loss in individuals over 65 in the western world. Although the mechanisms and progression of AMD are not fully understood, drusen and hyperreflective foci are among the key earliest signs of the disease and monitoring them using spectral domain optical coherence tomography (SD-OCT) imaging can help clinicians improve patient stratification and define the risk of disease’s progression. Observing patients over time with multiple assessments, as opposed to a single observation, increases the success in identifying the risk of progression. Also, the subjective task of assessment of huge amounts of imaging data in a clinical setting by hand is a problem that only recently the advances in computer-based image analysis have begun to solve. Moreover, computerized methods have opened the horizon of the search for under-recognized image biomarkers of AMD progression. We created tools and incorporated them in a unified graphical user interface (GUI) to automatically quantify and analyse drusen evolution over time in SD-OCT images of dry AMD patients using classic approaches and the App Designer functionality of MATLAB. The algorithm uses an improved segmentation of Bruch's membrane provided by the SPECTRALIS software to align the SD OCT image volumes relative to this layer. Then, within a user-defined central square around the fovea, it identifies and segments cuticular and soft drusen with a combined approach of two methods: a topographic analysis on slices near the maximum height of drusen and segmenting the RPE inner boundary (iRPE) in four different planes of the aligned SD-OCT volume. The watershed method is used to segregate drusen. Moreover, the algorithm automatically computes the drusen features relative to the overall region of interest (number of drusen, the percentage of drusen area in space within the user defined central region of interest, and the percentage of drusen volume within the same space, relative to the volume from Bruch’s membrane to the outer nuclear layer) and also per drusen (area of base, volume, median homogeneity inside and percentage of the presence of hyperreflective foci in all drusen identified). Finally, the algorithm can track the evolution of all of these features as the SPECTRALIS software performs intra-patient image alignment, although, in the future, it would be good to add a verification step to confirm they are properly aligned. Also, to increase usability, the GUI lets the user edit the segmentation of BM employed to align the images and to semi-manually revise the identified and segmented drusen. These algorithms were validated by comparing their performance with identifications performed manually using the GUI developed on a small sample of 10 patients and, using the algorithm in its entirety, achieved median sensitivity and precision of 0.88 and 0.92 respectively. The algorithm developed to automatically segment drusen also achieved medians accuracy, precision, sensitivity, and specificity of 0.92, 0.8, 0.9, and 0.92 respectively when validated on the same patients by evaluating its effectiveness in assessing the percentage of the total area occupied by drusen. Additionally, we computed the characteristics of the drusen of images of the 10 individual acquisitions used in the validation and also of images of 6 patients with 2 acquisitions spaced in time between 12 and 19 months. It proved the usability and good performance of all tools of the GUI, including the automatic longitudinal analysis tool. Nevertheless, both algorithms should be compared with identifications and segmentations performed by experts and tested in a higher number of patients. Also, the algorithms developed to perform volumetric segmentation and to obtain the drusen-specific features still need validation from specialists. Our results demonstrate the potential of the developed tools and GUI in a clinical setting to track the progression of dry AMD in patients and discover new biomarkers. It does not require a powerful processor while having a low processing time, making it easier to implement in the current clinical settings. At the same time, in the future, more tools can be incorporated such as a disease prediction algorithm regarding the established features and an interactive dashboard displaying the tracking of the features