Dynamic Inductive Transfer Learning with Decision Support Feedback to Optimize Retina Analysis

Muhammad Zubair Khan, Yugyung Lee

Abstract: The retina is a unique tissue considered an extension of a brain that transforms the incoming light into neural signals. Numerous deep neural networks are developed to segment retinal images precisely for detecting diabetic retinopathy and glaucoma. However, these networks are limited in selecting evaluation metrics and tuning hyperparameters subjectively in the model validation process. Furthermore, the segmentation networks lack a progressive mode of model tuning for active transfer learning. This article proposed a novel technique of dynamic inductive learning with single-point decision criteria, striving to optimize the image segmentation model using multi-criteria decision support feedback. A case study is conducted to reveal the problems related to the conventional approach and establish the significance of a proposed concept with empirical evidence. It is found that dynamic inductive transfer learning reduces the subjectivity of hyperparameter selection in a model validation process. For a given challenge of retinal vessel extraction, stochastic gradient descent is considered an optimal candidate for two variants of dynamic inductive transfer learning with a decision score of 0.9634 and 0.9951, respectively. This effort would serve as a vital step towards an optimal disease diagnosis.