Adaptive Learning Rate Adjustment via Type-2 Fuzzy Logic in Convolutional Neural Networks for Diabetic Retinopathy Detection and Classification

Abstract

One problem that the medical field hasfaced is the early detection of various existing diseases.Patients with diabetes mellitus are prone to additionalconditions, one of which is diabetic retinopathy. Due tothe increasing number of people with diabetes mellitus,the number of expert technicians is insufficient toadequately treat them. To solve this problem, computertools have been used to automate the detection ofdiseases. One of these tools is the use of artificialneural networks. These networks have the characteristicthat they can be adapted to a specific disease, whichallows for the creation of different neural network models.Each model has parameters that adjust the weights ofits neurons. These parameters are assigned by thedesigner of network architecture before training. Thisrequires time for testing and fine-tuning the parametersuntil the desired result is obtained. One of theseparameters is the learning rate of the training algorithm.This value can only be modified before training, soselecting the most appropriate one may require asignificant investment in time and analysis. This workproposes a method that adjusts the learning rate of theAdaptive Moment Estimation training algorithm betweeneach epoch using interval Type-2 and generalizedType-2 fuzzy inference systems, taking as input theaverage training and validation loss values, as well asthe epoch number. This reduces analysis time, allowingfor a focus on other network parameters. The proposedmethod is applied to two different convolutional neuralnetwork architectures: disease detection and classifyingdisease severity.