Ocular diseases are one of the main causes of irreversible inability of persons in productive age. In Colombia and other developing countries, due to the low coverage of public health systems, the lack of medical specialists and the high cost of specialized medical exams, less than half of the patients are correctly diagnosed. The diagnosis of these diseases heavily relies on medical diagnostic images, such as eye fundus images, angiography and optical coherence tomography which are acquired using specialized equipment. Several recent research works have shown the feasibility and utility of automatic analysis of these images to support diagnostic processes. In addition, the development of multimodal learning models has shown that the combination of different sources of information helps in retrieval and classifications tasks related to medical images. This research aims to design and implement multimodal learning systems for the processing of medical diagnostic images and medical records, oriented to detection and classification of eye pathologies. These multimodal systems imply in turn challenges of learning representation and fusion mechanisms, and have a variety of applications beyond the classification problems, which will help to validate the obtained results. 

The world of ophthalmology has been transformed by our ability to image the fundus of the eye. The problem is that even with the best clinical cameras, by the time changes indicative of disease are detected, hundreds of thousands of retinal cells have already been lost. Technology now exists to visualize individual cells in the living eye. However, for this to have a clinical application, we need robust tools that can identify the cells and make quantitative measurements on them. Although methods based on neural networks and deep learning have shown great promise in providing this analytical capability, these studies have been limited in scope to the analysis of specific diseases and the localization of lesions at the retinal level rather than at the cellular level. Our goal is to develop a machine learning model based on deep neural networks to generate robust cone identifications in retinal adaptive optics images, enabling quantitative analysis at the cellular level.