Volume 15, Issue 6 And S7 (Artificial Intelligence 2025)
J Research Health 2025, 15(6 And S7): 723-744 |
Back to browse issues page
Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
Chellapandi A, Rengaraj A, Kaliannan S. Analyzing the Model Performances of Dot Hemorrhage Pattern Recognition Using Deep Neural Networks. J Research Health 2025; 15 (6) :723-744
URL: http://jrh.gmu.ac.ir/article-1-2889-en.html
URL: http://jrh.gmu.ac.ir/article-1-2889-en.html
1- Department of Electronics and Communication Engineering, SRM Institute of Science and Technology Ramapuram Campus, Tamilnadu, India.
2- Department of Electronics and Communication Engineering, SRM Institute of Science and Technology Ramapuram Campus, Tamilnadu, India. ,arthir2@srmist.edu.in
2- Department of Electronics and Communication Engineering, SRM Institute of Science and Technology Ramapuram Campus, Tamilnadu, India. ,
Abstract: (1744 Views)
Background: Diabetic retinopathy (DR) is a leading cause of blindness, and DR progression can be avoided by early dot hemorrhage (DH) detection. Manual diagnosis is frequently subjective and time-consuming, which emphasizes the necessity for automated methods.
Methods: Using machine learning (ML) and artificial intelligence (AI), this study suggests an automated DH detection system. After preprocessing retinal fundus images with contrast limited adaptive histogram equalization (CLAHE), a modified VGG-16 convolutional neural network (CNN) was used to extract features. Support vector machine (SVM), random forest (RF), and linear regression (LR) models were used to classify both local and global features.
Results: The proposed approach is highly effective in detecting and classifying DH associated with DR. When tested on the APTOS dataset, the model obtained an overall accuracy of 93.26% for DH identification and classification. This shows that the CNN is extremely effective in learning the distinguishing characteristics of hemorrhagic lesions from retinal fundus images. To confirm the suggested approach’s efficiency, comparative research was carried out using three traditional ML algorithms: RF, SVM, and LR. Among them, the RF classifier had the greatest accuracy of 92.4%, surpassing both SVM and LR.
Conclusion: The research introduces a VGG16-based CNN that detects retinal hemorrhages in the APTOS dataset with an accuracy of 93.26%. Experimental results show that the Leaky ReLU activation function improves image classification performance, whereas the Adam and Adadelta optimizers consistently improve CNN-based learning. Finally, the research showed that deep learning-driven dynamic image analysis is a reliable and effective method for automated identification and categorization of DH, paving the way for early and accurate DR screening systems.
Methods: Using machine learning (ML) and artificial intelligence (AI), this study suggests an automated DH detection system. After preprocessing retinal fundus images with contrast limited adaptive histogram equalization (CLAHE), a modified VGG-16 convolutional neural network (CNN) was used to extract features. Support vector machine (SVM), random forest (RF), and linear regression (LR) models were used to classify both local and global features.
Results: The proposed approach is highly effective in detecting and classifying DH associated with DR. When tested on the APTOS dataset, the model obtained an overall accuracy of 93.26% for DH identification and classification. This shows that the CNN is extremely effective in learning the distinguishing characteristics of hemorrhagic lesions from retinal fundus images. To confirm the suggested approach’s efficiency, comparative research was carried out using three traditional ML algorithms: RF, SVM, and LR. Among them, the RF classifier had the greatest accuracy of 92.4%, surpassing both SVM and LR.
Conclusion: The research introduces a VGG16-based CNN that detects retinal hemorrhages in the APTOS dataset with an accuracy of 93.26%. Experimental results show that the Leaky ReLU activation function improves image classification performance, whereas the Adam and Adadelta optimizers consistently improve CNN-based learning. Finally, the research showed that deep learning-driven dynamic image analysis is a reliable and effective method for automated identification and categorization of DH, paving the way for early and accurate DR screening systems.
Keywords: Dot hemorrhages (DH), Convolutional neural networks (CNN), Deep learning (DL), Machine learning (ML), Contrast limited adaptive histogram equalization (CLAHE)
Type of Study: Orginal Article |
Subject:
● Artificial Intelligence
Received: 2025/08/22 | Accepted: 2025/11/1 | Published: 2025/12/31
Received: 2025/08/22 | Accepted: 2025/11/1 | Published: 2025/12/31
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |
This is an open access article distributed under the terms of the Creative Commons Attribution License (CC-By-NC), which permits use, distribution, and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes.
Contact Information
