Towards robust deep learning for medical imaging Applications in ophthalmology and radiology

Deep learning (DL) models for medical image analysis have gained substantial popularity over the past decade. Despite high performance in lab settings, these models often fall short when applied in the real world. This can have considerable negative impact on both trust in these algorithms and their clinical use. A major cause of this failure is the absence of intrinsic mechanisms within most systems to enhance robustness. In this thesis, we addressed three facets of robustness: (1) the model’s ability to perform well on data very close to or identical to the training distribution, (2) the model’s ability to flag ungradable images or out-of-distribution data, and (3) the model’s ability to maintain high performance when data in the deployment setting is slightly or locally unaligned with the training distribution. We developed and validated approaches to address these robustness facets in radiology and ophthalmology applications. For radiology, we focused on the rapid triage of pulmonary coronavirus disease 2019 using chest computed tomography scans, employing automated disease likelihood assessment and severity scoring. For ophthalmology, we devised and evaluated methods for increased robustness of DL systems for various retinal diseases and imaging modalities. These approaches were applied to systems for glaucoma screening using color fundus photographs, along with models for staging and biomarker quantification of age-related macular degeneration from optical coherence tomography scans.