Deep Learning Techniques for Automated Analysis and Processing of High Resolution Medical Imaging

Abstract

Medical imaging plays a prominent role in modern clinical practice for numerous medical specialties. For instance, in ophthalmology, different imaging techniques are commonly used to visualize and study the eye fundus. In this context, automated image analysis methods are key towards facilitating the early diagnosis and adequate treatment of several diseases. Nowadays, deep learning algorithms have already demonstrated a remarkable performance for different image analysis tasks. However, these approaches typically require large amounts of annotated data for the training of deep neural networks. This complicates the adoption of deep learning approaches, especially in areas where large scale annotated datasets are harder to obtain, such as in medical imaging. This thesis aims to explore novel approaches for the automated analysis of medical images, particularly in ophthalmology. In this regard, the main focus is on the development of novel deep learning-based approaches that do not require large amounts of annotated training data and can be applied to high resolution images. For that purpose, we have presented a novel paradigm that allows to take advantage of unlabeled complementary image modalities for the training of deep neural networks. Additionally, we have also developed novel approaches for the detailed analysis of eye fundus images. In that regard, this thesis explores the analysis of relevant retinal structures as well as the diagnosis of different retinal diseases. In general, the developed algorithms provide satisfactory results for the analysis of the eye fundus, even when limited annotated training data is available.