An active deep learning method for diabetic retinopathy detection in segmented fundus images using artificial bee colony algorithm

• Published in: The Artificial intelligence review Vol. 56; no. 4; pp. 3291 - 3318
• Main Author: Özbay, Erdal
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 01.04.2023; Springer; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Artificial neural networks; Blindness; Computer Science; Datasets; Deep learning; Diabetes; Diabetic retinopathy; Diabetics; Equipment and supplies; Feature extraction; Image analysis; Image processing; Lesions; Machine learning; Methods; Multilayers; Neural networks; Ophthalmology; Performance evaluation; Retinal images; Search algorithms; Swarm intelligence; Type 2 diabetes; Diabetic retinopathy; Image segmentation; Artificial bee colony; Active deep learning; Convolutional neural network
• ISSN: 0269-2821; 1573-7462
• DOI: 10.1007/s10462-022-10231-3

Retinal fundus image analysis (RFIA) is frequently used in diabetic retinopathy (DR) scans to determine the risk of blindness in diabetic patients. Ophthalmologists receive support from various RFIA programs to cope with the detection of visual impairments. In this article, active deep learning (ADL) using new multi-layer architecture for automatic recognition of DR stages is presented. In order to facilitate the detection of retinal lesions in the ADL system preprocessing, the image is segmented using the artificial bee colony (ABC) algorithm with a threshold value determined according to the results of the image histogram. Besides, a tag-efficient convolutional neural networks (CNN) architecture known as ADL-CNN has been developed to automatically extract segmented retinal features. This model has a two-stage process. In the first, images are selected to learn simple or complex retinal features using basic accuracy labels in the training examples. Second, useful masks are provided with key lesion features and segment areas of interest within the retinal image. Performance evaluation of the proposed ADL-CNN model is made by comparing the most advanced methods using the same dataset. The efficiency of the system is made by measuring statistical metrics such as classification accuracy (ACC), sensitivity (SE), specificity (SP), and F-measure. The ADL-CNN model applied to the EyePacs dataset containing 35,122 retinal images yielded 99.66% ACC, 93.76% SE, 96.71% SP, and 94.58% F-measure. In this respect, it can be said that the proposed method shows high performance in detecting DR lesions from various fundus images and determining the severity level.