Multi Scale aided Deep Learning model for High F1-score classification of fundus images based Diabetic Retinopathy and Glaucoma

Diabetic Retinopathy (DR) and Glaucoma (GL) are leading causes of irreversible visual loss in individuals aged 30-45 years due to optic nerve and retinal blood vessel damage. Fundus images of the retina are used to investigate DR & GL, but manual screening is error-prone, laborious, and tedious. Capturing fundus images with different fields of view and resolutions may result in the loss of important features, leading to misclassification. This paper proposes a multi-scale aided Deep Learning (DL) model for accurate DR & GL classification. To reduce the computational burden, an Atrous Spatial Pyramid Pooling (ASPP) layer is introduced to transfer features to the DL model-based classifier. Precision, Recall and accuracy are improved by incorporating Soft F1-Loss and Categorical Cross Entropy (CCE) Loss in the modified Loss function, enhancing the F1-score and Accuracy. The proposed model is validated using different DL models on a dataset of 10,500 images from different public and private databases. The results show that the Multi-scale RESNET50 model achieves 87.5% accuracy and 0.89 F1-score, outperforming the Multi-scale Deep Convolutional Neural Network (CNN) model. This approach offers a promising solution for accurate DR & GL detection, facilitating early diagnosis and intervention in these sight-threatening conditions.