Thesis defense Cristina González Gonzalo (Donders Series 630)
18 December 2023
Promotors: Prof. dr. ir. C.I. Sánchez Gutiérrez (University of Amsterdam), Prof. dr. B. van Ginneken
Trustworthy AI for automated screening of retinal diseases
Screening for retinal diseases has become a high-priority healthcare service to prevent vision loss. When available, screening protocols rely on periodical, manual examinations of the retina by highly specialized workforce. This fails to meet the requirements of large-scale screening, especially in low-resource countries. The potential of artificial intelligence (AI) to automate the screening of retinal images has become apparent in recent years. Deep learning (DL) systems have been shown to achieve performances close or even superior to that of experts for the detection of high-prevalence diseases, such as diabetic retinopathy and age-related macular degeneration. However, few AI systems count with regulatory approval for being used in real-world screening settings.
This thesis focuses on the development of algorithms and the study of different factors that contribute to enlarge the gap between AI development and its integration in ophthalmic practice. We investigate the reliability of commercially-available software for automated screening of retinal diseases; the explainability of DL systems’ decisions and its impact on trust and clinical usability; the robustness of DL systems against malicious attacks; and how to generate trustworthy AI, the next step necessary to bring the benefits of AI closer to the final users in healthcare.