MSM2: Multi Scale and Multi Model based Diabetic Retinopathy detection with TTA

Diabetic Retinopathy has become a significant problem across the globe, causing blindness, due to which its identification has gained prime importance so that it can be cured at early stages. Many researchers have proposed machine learning and deep learning-based methods, but there is still room for improvement. So, we propose a method for its detection which uses multi-scale ensembling, global weighted average pooling (GWAP) with Dihedral TTA(Test Time Augment), which helps to improve the results. Additionally, we have employed a range of augmentations to enhance generalization and utilized oversampling to address class imbalance on the APTOS 2019 dataset. Our approach demonstrates significant improvement across all evaluation metrics. Using Dihedral TTA helped boost precision, recall, and F1 score compared to normal inference without dihedral TTA, resulting in more stable and reliable DR severity predictions.