STMFNet: Spatial Texture Multi-Scale Feature Fusion Attention Network for Diabetic Retinopathy Classification

Diabetic Retinopathy (DR) is a major cause of vision loss and blindness, particularly among diabetic patients. Effective and timely treatment of DR relies on precise and automated detection systems that can assess disease severity from retinal fundus images. Traditional clinical approaches are often time-consuming, and earlier texture attention models may struggle to accurately detect subtle features, such as microaneurysms or abnormal blood vessel patterns, which are crucial for early diagnosis. To address this, we proposed the STMFNet model, a hierarchical framework designed for classifying 11 stages of DR severity. The model combines two primary mechanisms: The Texture Spatial Attention Network, which focuses on identifying critical texture features related to DR while minimizing irrelevant background information through attention gating, and the EfficientNet backbone with Multi-Scale Feature Fusion, which captures a wide range of image patterns. By extracting and fusing features from different layers of EfficientNet-V1 BO, the model effectively learns both low (e.g., edges, blobs) and high (e.g., objects, patterns) level representations. These features are further refined through spatial multi-scale attention, and the final classification into 11 DR stages is achieved through a Fully Connected Network (FCN) and SoftMax. Our experimental results show that STMFNet significantly outperforms existing state-of-the-art models on the publicly available Kaggle fundus dataset, demonstrating its potential for reliable DR diagnosis in clinical settings.