Improved GoogLeNet based on CBAM module for fundus disease classification network

Fundus images are a valuable tool in the diagnosis of ophthalmic and systemic diseases. However, the traditional manual analysis process is time-consuming, labour-intensive and susceptible to subjective factors, which presents a challenge in the context of efficient and accurate diagnosis. It is therefore imperative to develop efficient and automated classification methods to enhance the accuracy and efficiency of diagnosis. In this experiment, a disease classification method combining multi-scale feature fusion with a spatial-channel attention mechanism (CBAM) is proposed for the automated classification of fundus images. The method employs GoogLeNet as the underlying network and utilises the Inception module to facilitate parallel extraction of multi-scale features, thereby enabling the effective capture of multi-level feature information. Furthermore, it is combined with the CBAM module, which enables the dynamic adjustment of weights for feature maps in both the channel and spatial dimensions. This significantly enhances the network's capacity to recognise disease features. The experimental results demonstrate that the enhanced model attains a classification accuracy of 92.64% on the publicly accessible OIA-ODIR dataset. This represents a notable enhancement in terms of evaluation metrics, including accuracy, precision, and F1 score, in comparison to the original GoogLeNet, VGG19, and ResNet50 models. Although the recall rate was slightly lower than that of ResNet50, the overall performance was excellent, thereby validating the effectiveness of the method in fundus disease classification and providing strong support for automated fundus disease diagnosis.