Intelligent classification of retinal lesions, it is difficult to get the right features. There are large differences in lesion areas and it is easy for fine structural features to be lost during this process, so we propose a deep learning model that combines multi-scale feature extraction and channel attention mechanism for these key problems in intelligent classification of retinal lesions. The model first designs a multi-scale feature fusion module, under different receptive field to get the lesions feature with parallel dilated convolutions having different dilation rates; secondly, it introduces an efficient channel attention mechanism that can adaptively re-calibrate the response weights of each feature channel and suppress irrelevant noise interference; finally, a joint loss function is used to optimize the model parameters. An empirical analysis was conducted on a large-scale public dataset of retinal lesions containing 35,126 fundus images, divided into training set, validation set, and test set according to the ratio of 8:1:1. Experimental results show that the model achieves a total classification accuracy of 97.83% on the test set, achieving an AUC value of 0.992, which is 4.21 percentage points higher than the baseline model ResNet50. The ablation experiments have proved the effectiveness of the multi-scale feature fusion and channel attention mechanism. It offers a viable technical option for automated screening for retinal issues.
