Deep hyperparameter transfer learning for diabetic retinopathy classification

The detection of diabetic retinopathy (DR) in millions patients across the globe is a challenging problem. Diagnosis lengthy and tedious process, requiring medical professional to assess individual fundus images patient's retina. This process can be automated by applying deep learning (DL) technology given huge dataset. problems associated with DL are unavailability large dataset their higher training time. model's best performance achieved using set optimal hyperparameters (OHPs) obtained performing costly iterations hyperparameter optimization (HPO). These addressed transfer (TL) technique both model HPO. TL HP tuning focus this work. authors study applicability EyePACS DR dataset's OHPs other datasets, forming basis research question classification performed ResNet trained on (kaggle) Indian image (IDRiD) datasets. Various HPs tuned work data augmentation configuration, number layers, optimizers, samplers, rate, momentum. demonstrate that suitable for IDRiD without needing tune from scratch. task reusability poorly reported literature. Therefore, here used researchers. Moreover, researchers working datasets also apply same since they reusable no HPO required. provided EyePAC after being scratch, which as starting point others.