Detecting cognitive impairment in diabetics based on retinal photos by a deep learning method

impairment in diabetic patients has drawn increasing attention, yet conventional assessments like neuroimaging and cognitive scales are costly, invasive, or subjective, limiting their use in large-scale screening. This study proposes a deep learning-based method for identifying moderate to severe cognitive impairment in type 2 diabetes patients using only fundus images. A total of 1,000 fundus images from 250 patients were collected. We developed a four-branch model, FB_Net, incorporating a self-designed Average Attention Block (AA_Block) and Multi-Scale Convolutional Block Attention Module (MS_CBAM). The latter introduced a Multi-Scale Convolution Block (MSC_Block) to enhance the multi-scale feature extraction capability of the original Convolutional Block Attention Module (CBAM). We compare four backbone networks—MobileNetV1, AlexNet, EfficientNet-b0, and ResNet34, among which MobileNetV1 achieved the best performance for classification, with an accuracy of 0.732 and an AUC of 0.790. Grad-CAM visualization revealed that regions rich in fundus vasculature are key to classification as biomarkers. These results highlight the importance of vascular features in cognitive assessment and demonstrate that the proposed artificial intelligence approach is a promising, noninvasive, and cost-effective tool for early screening and potential clinical application in diabetic populations.