Abstract
Deep learning and artificial neural networks have been extensively applied to the automated diagnosis of retinal diseases from fundus images. Recent advancements have also led researchers to leverage deep learning to examine the connections between the retina and systemic health in a discipline termed oculomics. However, as oculomics models likely combine multiple retinal features to arrive at their conclusions, traditional methods in model interpretation, such as attribution saliency maps, often provide uncompelling and open-ended explanations that are prone to interpretation bias, highlighting a need for the examination of alternative strategies that can quantitatively describe model behavior. One potential solution is neuron activation patterns, which were previously applied to real-time fault diagnosis of deep learning models. In this study, we proposed a novel and experimental framework of neuron activation pattern synthesis leveraging image similarity metrics, with the outcome being a continuous, metric-based descriptor of underlying model behavior. We applied our approach in examining a model predicting systolic blood pressure from fundus images trained on the United Kingdom Biobank dataset. Our results show that the metric-based descriptor was meaningfully related to cardiovascular risk, a real-life outcome that can be expected to be related to blood pressure-related biomarkers identified from a fundus image. Furthermore, it was also able to uncover two biologically distinct and statistically significant groups among participants who were assigned the same predicted outcome and whose distinctness would otherwise be imperceivable without the insights generated by our approach. These results demonstrate the feasibility of this prototypical approach in neuron activation pattern synthesis for oculomics models. Further work is now required to validate these results on external datasets.