What is it about?
Biomedical Imaging and Deep Learning(DL / AI) works have established their positions, but how well they actually fit together. This paper examines and exhibits a ray of hope that diabetic retinopathy detection with deep learning methods can change the whole existing scenario in the direction of providing confidence to the medical experts to detect and treat better and efficiently.
Featured Image
Photo by Amanda Dalbjörn on Unsplash
Why is it important?
We used a tripartite random selection procedure to split the dataset into three groups: training (84 participants), validation (20 participants), and testing (30 participants). The suggested technique underwent 100 training rounds on a server with 11 GB of RAM and an NVIDIA 1080Ti GPU. The entire time for the race was twelve hours. Using Adam Optimizer and a 0.0001 learning rate, we trained the model. We used the ReLU activation function and applied batch normalization. By standardizing the network at each layer, batch normalization often makes the model more stable. There are quantitative and qualitative assessments of the suggested approach.
Perspectives
One of the many long-term effects of diabetes is DR. The five phases of DR, which may vary from moderate to severe, are a result of the intensity variation. In this study, we used a hybrid method to detect several DR lesions in the retina without human intervention. Images of the Messidor fundus found in public databases serve as both the model's training and testing grounds. The suggested network accurately identifies the DR phases according to the experimental results. On pictures of healthy retina, stage 1, stage 2, and stage 3 fundus, the suggested approach achieves a sensitivity of 93.54%, a precision of 91.34, an F1-score of 90.34, and a loss of 0.423. Our suggested network performs competitively when compared to other techniques. The results demonstrated the feasibility of implementing our method for CFPs analysis into routine clinical practice and provide credence to the idea that screening programs should be thorough in order to detect DRs at an early stage. Since there are not yet many publicly available big training fundus image datasets with several lesion labels to support multi-label training, researchers will have to settle with doing more trials on bigger datasets in the future in order to improve prediction accuracy.
Agha Urfi Mirza
University Of Technology And Applied Sciences, Al Musana, Oman
Read the Original
This page is a summary of: Diabetic retinopathy detection using deep learning approach, January 2025, American Institute of Physics,
DOI: 10.1063/5.0275788.
You can read the full text:
Resources
Contributors
The following have contributed to this page
