Automatic diagnosis of eyelid tumors based on SE-SSD object localization algorithm

Eyelid tumors include tumors of orbit, conjunctiva, tissues of the eyeball and eye appendages, which seriously affect people's vision and health. The similarity of early benign and malignant structures makes it difficult for ophthalmologists lacking clinical experience to distinguish them. To address these issues, a target localization algorithm based on SE-SSD (ResNet50) is proposed, which integrates the channel attention mechanism SENet into the SSD algorithm to realize automatic diagnosis of benign and malignant eyelid tumors. Specifically, we utilize the SE-SSD (ResNet50) network to locate the lesion regions of eyelid tumors, and subsequently employ DenseNet121 to achieve benign and malignant classification of eyelid tumors. Experimental results demonstrate that compared with the SSD algorithm, the AP of SE-SSD (ResNet50) is increases by 0.029 on the test set. In the experiment for benign and malignant eyelid tumor classification, DenseNet121 achieves an AUC of 0.919 (95% CI: 0.863-0.961) and an accuracy of 90.9% (95% CI: 86.9-95.0) on the test set. The experimental results indicate that our model has excellent performance in the automatic diagnosis of benign and malignant eyelid tumors.