An automated diabetic retinopathy of severity grade classification using transfer learning and fine-tuning for fundus images

Abstract

Diabetes mellitus Retinopathy (DR) has recently become a major health problem, and its complications are also increasing worldwide. Early diagnosis of DR is essential to determine the significance of several features from fundus images for detection and classification in many Computer-Aided Diagnosis (CAD) applications. However, existing methods suffer from high dimensional features, small training datasets, misclassification, and high training loss, which leads to a complex grading system. Aiming at these concerns, this paper presents a Frame-wise Severity Scale Classification Model (FSSCM) using Transfer Learning enabled EfficientNet B3 and Fine Tuning enabled ResNet 101, namely, TL-EN3 and FT-RN 101, to classify the severity of disease level of retinal fundus images. Initially, the preprocessing and augmentation processes are performed to bring out the clear view features of the raw fundus images. Then the segmentation phase constrains the whole region using the Chan-Vese algorithm. Twelve features are extracted and fed into the learning network for training purposes. The proposed work utilizes the TL-EN3 model to capture high-resolution patterns with high accuracy and integrates FT-RN 101 models to maintain a balance between efficiency and accuracy with fewer parameters. Experimental analysis is conducted with different metrics such as kappa coefficient (K-score), classification accuracy (CA), precision (P), recall (R), F1-measure (F1), and False Positive Rate (FPR) on three publically available datasets such as Kaggle, Messidor-1, and Messidor-2 datasets. Furthermore, some performance graphs are plotted for visualizing the architecture performance, including training loss, validation loss, training accuracy, and validation accuracy. The performance of the proposed FSSCM approach obtains high estimation values of 0.981 0.985 0.983, 0.98 0.986 0.984, and 0.98 0.985 0.98 in terms of P, R, and F1 on three datasets, respectively. Also, it achieves high estimation results of 99.02 0.993, 98.1 0.97, and 98.3 0.98 in terms of CA and K-score for three datasets, respectively. With a high training accuracy and a low level of training loss, the proposed method gets better severity level classification results than other models.