Fundus Image Classification: A Wavelet Feature Descriptor Approach

Lately, many diabetic patients are experiencing diabetic retinopathy resulting in a loss of their sight. Even though the urgency and threat posed by this condition, there is insufficient data source to engage appropriate computational intelligence tools. The few that exist happen to be imbalanced. Leveraging on this imbalanced dataset, several activities have been carried out to propose improved detection and classification descriptors. Although some works have been done in this domain, the issue of accuracy still persists in the administration of an effective diagnosis. This paper harnessed the benefits of Gabor filters and the multi-resolution property of Discrete Wavelet Transforms (DWTs) to construct appropriate fundus feature descriptors. These discriminant features are fed into some selected but predominant classical machine learning classifiers. Numerical evaluation of the study gave a perfect (100
^%
) average score for the fundus image classification using Gradient Boosting and Logistic Regression classifiers over Accuracy, F1-score, Precision and Recall evaluation metric. The tie in performance is further broken using their computation time, suggesting that Logistic Regression is more appropriate with 9min 32sec over Gradient Boosting or 1hr 10min 32sec.