SEARCH RESULT

The Diabetic Retinopathy (DR) continues to be a leading cause of blindness worldwide, fueled by diabetic complications that lead to retinal damage. Early diagnosis via retinal fundus imaging is critical to avoid irreversible blindness. This process of manual grading of the images is time-consuming and vulnerable to human error. With advancements in machine learning (ML) and deep learning (DL), autonomous systems have proved capable of outperforming conventional diagnostic methods. This article reports a systematic review of latest developments in DR detection and classification using fundus images, comparing the performance of different ML and DL methods. It discusses fundamental aspects of the diagnostic pipeline, such as image pre-processing, data augmentation, feature extraction, and classification algorithms. The review also discusses the application of Federated Learning (FL) as a privacy-maintaining method for decentralized healthcare data. Benchmark datasets, evaluation metrics, and main challenges in clinical integration are addressed. The paper posits that the integration of DL architectures with secure learning algorithms such as FL can result in more efficient and scalable DR diagnostic systems, leading to improved clinical decision support.

International Conference on Advanced Technologies, Computer Engineering and Science
ICATCES

Jagruth K Raj Hemanth B M Jayanth R Rao Harini D K Koushik A R

Diabetic Retinopathy (DR) is a leading cause of visual loss and blindness in adults in their active working age. Early detection and accurate classification of retinal pathology are critical to enable informed clinical decisions. This paper suggests an end-to-end deep learningbased retinal disease segmentation and disease severity classification system to aid automatic diagnosis by an ophthalmologist. The new system will address the two most important problems of lesion segmentation and disease grading. For lesion segmentation, object-level and pixellevel approaches will be employed. The microaneurysms, hard exudates soft exudates, and hemorrhages will be segmented accordingly as signature lesions using a pretrained encoder-decoder model based on convolution with DeepLabV3 architecture. Segmentation will be boosted by adding YOLOv8, a new state-of-the-art object detection model that can do fast detection and localization of retinal lesions. For grading of disease severity, a single-task multioutput CNN classifier will identify the severity of Diabetic Retinopathy and Diabetic Macular Edema risk from retina fundus images. The classification model exhibited significant training accuracy and test data generalization. Finally, the entire pipeline is made available as a point-and-click web application based on Flask so that users can upload retinal images and obtain segmented lesion outputs along with real-time disease grade predictions.

International Conference on Advanced Technologies, Computer Engineering and Science
ICATCES

Jagruth K Raj Hemanth B M Jayanth R Rao Harini D K Koushik A R