Optimizing deep transfer networks with fruit fly optimization for accurate diagnosis of diabetic retinopathy

Abstract

It is crucial to develop a smart analytics system capable of accurately diagnosing diabetic retinopathy. This research uses a new deep transfer network framework to diagnose Diabetic Retinopathy (DR). The core of this framework is to employ a new Fruit Fly Optimization Algorithm (MALBFOA) enhanced by the Levy Flight (LF), Gaussian Transboundary Correction (GTC), Multi-subgroups, and Subgroups Annihilation (SA) mechanisms to optimize two fully connected layers parameters in one transfer deep learning model and establish a MALBFOA-based Deep Learning (MALBFOA-DL) for diagnosing diabetic retinopathy with a large set of color fundus photography obtained under a variety of imaging conditions as input. To verify the proposed method’s effectiveness, we quantitatively compare the proposed MALBFOA with the original FOA, FOA-based variants, and other traditional meta-heuristic algorithms in a comprehensive set of 49 benchmark functions (shifted and swirled). The experimental results validate that MALBFOA holds a faster convergence rate and better solutions in almost all benchmark functions, especially in solving asymmetric complicated optimization problems. The proposed MALBFOA-DL model can also grade the degree of diabetic retinopathy with more accurate recall rates than the benchmark model and assist doctors in diagnosing diabetic retinopathy.