Exploring Deep Learning Techniques for Supporting Eye Bank Operators in Polymegethism Assessment

Corneal transplantation is a practice that often constitutes the only viable option for patients with damaged corneas to recover vision and quality of life. In order to ensure safety of the operation, Eye Banks have the paramount role of assessing the donor corneas quality, through visual inspection and assessment of underlying conditions that could inhibit success of the operation. Among these conditions is polymegethism, i.e. the abnormal variation in size of endothelial cells, whose detection is often performed manually, requiring a strong investment in time and resources. Total or partial automation of this crucial step can lead to a drastic reduction in processing time, effectively speeding up the pipeline leading to corneal transplants, benefiting both patients in need and medical operators. This thesis project was developed during an internship at NIDEK Technologies Srl. in collaboration with FBOV (Fondazione Banca degli Occhi del Veneto), and revolves around exploiting deep learning techniques to explore partial automation of the polymegethism assessment process. The aim of the project is to explore the feasibility of developing a deep learning system that can fully or partially automate the process of estimating cell area variability from light microscopy images of the corneal endothelium, employing point annotations on cells positions. The project initially explores how annotations originally designed for cell counting can be repurposed for polymegethism detection. These annotations are leveraged to derive both interpretable spatial representations, in the form of visual maps, and scalar descriptors quantifying polymegethism. Multiple learning paradigms are investigated, including direct regression, dense prediction, and a two-stage approach combining both, all employing Convolutional Neural Networks to process raw light microscopy images. Results show high visual accuracy in the dense prediction task, with models effectively capturing meaningful spatial structures in the form of interpretable density and segmentation maps. Direct regression approaches exhibit notable limitations in representing target variability, partially mitigated by the two-stage approach, which leverages intermediate spatial representations to guide regression and improve performance. An intra-annotator variability analysis further contextualizes these findings, suggesting that model performance is constrained by the data itself, with the two-stage model approaching ceiling performance. This work establishes a foundation for the partial automation of polymegethism detection, demonstrating that existing annotations can yield meaningful qualitative outcomes despite not being specifically designed for this task. The intra-annotator variability findings further suggest, however, that a task- specific annotation strategy may be necessary to achieve reliable quantitative predictions.