Attention-Boosted Ensemble of Pre-Trained Convolutional Neural Networks for Accurate Diabetic Retinopathy Detection

Early and accurate detection of diabetic retinopathy (DR) is crucial for preserving vision. Extracting informative and global features that facilitate precise and reliable decision-making is essential. Convolutional neural networks (CNNs), known for their high accuracy, are well-suited for this application. However, these models are susceptible to data scarcity, a challenge that can be mitigated through transfer learning. Additionally, model ensembles have proven effective in similar domains. This study proposes the use of two pre-trained CNNs from the EfficientNet family, specifically EfficientNetB0 and EfficientNetB1, in conjunction, and combines the features from both models to enhance decision-making. A Multi-Head Attention layer is incorporated to extract global and region-independent features, further improving representation. Consequently, the model can focus on the most critical areas of the image, thereby increasing detection accuracy. The proposed approach is evaluated on two datasets, yielding impressive results in binary classification (DR or No-DR) on the IDRiD dataset, it achieved an accuracy of 99.07% and an F1 score of 99.02%, while on the APTOS dataset, it attained an accuracy of 99.19% and an F1 score of 99.07%. These findings illustrate the effectiveness of combining CNNs with attention mechanisms for the accurate and timely diagnosis of DR.