EfficientNet embedded with spatial attention for recognition of multi-label fundus disease from color fundus photographs

Abstract

Color fundus photographs enable the observation of numerous critical biomarkers and early-onset lesions associated with illnesses. Due to its non-invasive and cost-effective nature, this approach can be used for large-scale screening of fundus disorders. In recent years, most applications of data-driven deep learning approaches to fundus illnesses have been based on color fundus photographs. However, current screening methods for fundus illnesses are mainly focused on identifying particular diseases. The majority of published models for multi-disease recognition are ensembles of several binary classification networks, and an ensemble network consumes more resources than a single network. Additionally, many fundus illness recognition models make direct use of established networks for natural picture processing without performing structural optimization, which may reduce disease classification accuracy. To address these issues, we optimized the network structure to build a single multi-label fundus disease recognition model. Specifically, we began by selecting EfficientNet-B4 as a backbone network, then modified the output layer to create a multi-label recognition model. Based on this model, the spatial attention structure was extended to improve the network’s feature extraction capability. The focal loss was used to increase classification accuracy by mitigating the training effects of data imbalance. Our proposed model significantly improves recognition performance on a publicly available dataset compared to other baseline networks. When spatial attention and focal loss function are introduced into the backbone model, the baseline values of all relevant evaluation indicators are improved without significant increase in computational cost. Additionally, this article provides two error correction strategies for multi-label classification issues: mutual exclusion and super label space. When a more robust error correction strategy is employed with the model, F1 increases by 2.86% in comparison to the original performance of EfficientNet-B4.