Comparing Conventional and Deep Feature Models for Classifying Fundus Photography of Hemorrhages.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Diabetic retinopathy causes visual impairment through retinal abnormalities.
• Accurate detection of hemorrhages is crucial for effective treatment.
• Identifying hemorrhages near blood vessels or the retinal border presents a significant challenge.

Purpose of the Study:

• To develop and compare a novel hemorrhage detection method for diabetic retinopathy.
• To evaluate the effectiveness of conventional versus deep learning features for classification.
• To address the challenge of detecting hemorrhages in difficult locations within retinal images.

Main Methods:

• Image preprocessing included adaptive brightness adjustment and contrast enhancement.
• Hemorrhage localization utilized Gaussian matched filtering, entropy thresholding, and morphological operations.
• Segmentation employed a novel technique based on regional intensity variance, followed by feature extraction using conventional and deep models for support vector machine training.

Main Results:

• The developed method successfully identified and segmented hemorrhages, including those in challenging locations.
• Both conventional and deep features yielded promising classification results.
• Deep learning models demonstrated superior performance compared to conventional features in classifying diabetic retinopathy-related hemorrhages.

Conclusions:

• The proposed hemorrhage detection and segmentation technique is effective for analyzing retinal images.
• Deep learning approaches offer enhanced capabilities for classifying diabetic retinopathy features.
• This research highlights the potential of deep models to improve the diagnosis and management of diabetic retinopathy.