A deep learning based system for detecting abnormalities in the lower eyelid of smartphone images

In response to the clinical demand for screening lower eyelid abnormalities, this study proposes a smartphone image detection system based on an improved deep learning model, aiming to enhance detection efficiency and accuracy and optimize medical resource allocation. The study collected 1490 images of lower eyelids taken by Huawei smartphones (749 normal cases and 741 abnormal cases), and constructed a diverse dataset through preprocessing such as random cropping, geometric and color enhancement, and noise injection. In terms of model, the improved C-ResNet101 enhances the feature extraction ability of the model by introducing CBAM attention mechanism and feature pyramid network (FPN), and optimizes the training by combining hierarchical learning rate and dynamic cosine annealing strategy, which significantly improves the performance of the model in identifying eyelid anomalies. At the same time, the YOLOv8 model was used to realize the accurate detection of the eyelid region, and the gradient flow information was improved and enhanced by the C2f module. Experimental results show that C-ResNet101 achieved an accuracy of 92.62%, specificity of 95.95%, and F1 score of 0.9241 on the test set, outperforming models such as ResNet101, DenseNet, and ConvNeXt. This system enables convenient screening of lower eyelid abnormalities through smartphones, with an accuracy approaching that of experts, providing an efficient auxiliary tool for primary medical care and family health management. In the future, it is necessary to further expand the dataset size, adapt to the image quality of multiple models, and optimize the model's generalization ability to promote its application in large-scale screening of ophthalmic diseases.