Label Smoothing Loss with Dual-Stream Network Using Separable Convolutional Layers for Retinopathy Grading and Classification

Retinopathy detection based on deep learning methods is a challenging problem, especially the diabetic retinopathy (DR) brings so many technical complications in medical image processing. Recently, label smoothing regularization has proved to be a better option to improve the performance of deep learning models. Therefore, in this paper, we introduce a dual-stream multi-task learning model along with a novel weighted label smoothing regularization loss (WLSRL) to detect retinopathy. The proposed model uses a dual-stream network by incorporating separable and conventional convolutional neural networks to detect diabetic retinopathy. The model is designed to classify numerous retinal diseases on two different types of data. The data $\Delta_{1}$ is based on stereoscopic fundus photographs and $\Delta_{2}$ consists of OCT-based retinal images. The model is trained and tested on both data separately. We perform two classification tasks TF1, TF2 on $\Delta_{1}$ and TO1, TO2 on $\Delta_{2}$. The task TF1 is for the classification of fundus photographs as normal and abnormal, whereas TF2 is for DR grading. Similarly, the task TO1 classifies OCT-based images into four classes, whereas the task TO2 classifies images as normal and abnormal. The empirical results show that the model achieves competitive results for retinopathy classification and grading using multitask learning with WLSRL.