No Detail Found!

A SMOTE-Enhanced Hybrid VGG16–ResNet-50 Model for Automated Diabetic Retinopathy Detection

Author(s)
Muhammad Faris Hamdard University, Karachi, Pakistan, PK
Shahzaib Khalid Hamdard University, Karachi, Pakistan, PK
Muhammad Dilawar Khan Hamdard University, Karachi, Pakistan, PK
Mir Farooq Ali Marche Polytechnic University, Ancona, Italy, PK
Muhammad Mansoor Mughal Hamdard University, Karachi, PakistanUniversity of Houston, Texas, USA, PK
Tariq Javid Hamdard University, Karachi, Pakistan, PK
Abstract
Diabetic Retinopathy (DR) occupies the top space among the preventable causes of vision loss across the globe with people who are diabetic presenting a higher number of victims. It suggests a hybrid deep learning network that consists of the VGG16 and ResNet-50 architecture to improve classification of the severity of DR based on the retinal fundus image. To fit the data model, a balanced and preprocessed dataset was used by applying data augmentation and Synthetic Minority Over-sampling Technique (SMOTE). Training was performed on input images which were normalized to 512 x 512 pixels and carried out over 25 epochs with a batch size of 32. The suggested model attained mean accuracy, precise, recall at 86%, 85%, 84% and F1- score at 85% respectively as compared to benchmark meaning that the model is capable of robust classification. Quantization-aware training was also used to maximize the computational efficiency of the model where the model now takes 95 milliseconds on average to process a single image, suitable to be deployed in a near real-time fashion (low resources). The hybrid model shows scalability with promise of inclusion in automated DR screening systems and will, therefore, provide an early solution to accurate diagnosis despite a few misclassifications that occurred due to the visual similarities between the DR stages.
Publication Details
Page(s) 1-10
DOI DOI not available
Published Journal: Allied Medical Research Journal, Volume: 3, Issue: 3, Year: 2025

Request the full text from PASTIC

Keywords
Diabetic Retinopathy, Fundus Image Classification, hybrid model, Quantization-aware Training, deep learning
References
References section not found
0

Citations

0

Downloads

12

Views