Classification of Retinal Fundus Images into Normal / Diabetic with Fused Deep Features

The global incidence of diabetes is steadily increasing, necessitating timely identification and intervention to reduce its impact. Diabetic-Retinopathy (DR) is a prevalent complication of diabetes and a primary cause of global blindness. Timely identification and intervention are crucial in order to regulate his severity. The detection of diabetic retinopathy (DR) at the clinical level is frequently conducted using image-supported approaches. The present study introduces a deep-learning (DL) technique designed for the automated identification of diabetic retinopathy (DR) by utilizing retinal fundus images (FI). The research procedure involves several stages: collecting retinal FI and performing initial processing, extracting features using selected deep learning algorithms, reducing features and combining them, and performing bi-level classification and confirming performance. This study examines the DenseNet121 (DN121) scheme and conducts bi-level classification using the SoftMax and other selected classifiers. The findings of this study validate that the DN121 model achieves classification accuracy over 91% when using individual features with SoftMax, and surpasses 98% when utilizing fused features with the Decision Tree classifier. This study yielded a noteworthy outcome, and the suggested deep learning method can be regarded as a viable option for assessing the clinical grade of retinal fibrosis in the future.