Automated Diabetic Retinopathy Detection Using Horizontal and Vertical Patch Division-Based Pre-Trained DenseNET with...

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Diabetic retinopathy (DR) is a leading cause of preventable vision loss, necessitating accurate and timely diagnosis.
• Manual interpretation of fundus photographs for DR detection is prone to human error and delays.
• Automated classification systems are crucial for efficient and reliable DR screening.

Purpose of the Study:

• To develop and validate a novel automated method for classifying diabetic retinopathy (DR) using fundus images.
• To enhance the accuracy and robustness of DR classification compared to existing methods.
• To introduce a new, non-fixed-size patch division model for improved feature extraction.

Main Methods:

• A patch-based deep-feature engineering model utilizing DenseNet201 architecture and transfer learning.
• Extraction of deep features from horizontal and vertical image patches.
• Feature selection using neighborhood component analysis and classification via a cubic support vector machine.
• Validation on a newly collected three-class DR dataset and the APTOS 2019 five-class dataset.

Main Results:

• Achieved 94.06% accuracy for three-class classification on the new dataset (80:20 hold-out).
• Attained 87.43% accuracy for five-class classification on the APTOS 2019 dataset (80:20 hold-out).
• Outperformed previous studies on the APTOS 2019 dataset by over 2% in classification accuracy.

Conclusions:

• The proposed patch-based deep-feature engineering model demonstrates high accuracy and robustness for automated DR classification.
• This cognitive model offers an efficient and effective approach for early detection and management of diabetic retinopathy.
• The method holds significant potential for improving DR screening programs and reducing vision impairment globally.