基于改进Xception网络的糖尿病视网膜病变分级研究OA

Objective To propose a deep learning model based on improved Xception network to enhance the accuracy of automatic grading for diabetic retinopathy(DR).Methods Xception was used as the backbone feature extraction network,which was composed of three stages of entry flow,middle flow and exit flow,and the multi-scale attention fusion module(MAFM)was introduced to construct a MAFM-Xception model.The MAFM consisted of two components of multi-scale feature pyramid(MFP)and dual attention mechanism(DAM),which effectively fused shallow-level detail information with deep-level semantic features by the MFP and adaptively enhanced key lesion regions in both the channel and spatial dimen-sions with the DAM,thereby improving the model's ability to comprehensively characterize multi-scale lesions.During model training,a class weighting strategy was used to mitigate data imbalance,and contrast-limited adaptive histogram equalization(CLAHE)was employed to enhance image contrast and aid feature learning.To validate the model's accuracy for DR auto-matic grading,ablation and comparison experiments were conducted on the Asia Pacific Tele-Ophthalmology Society 2019(APTOS-2019)blindness detection dataset.The accuracy and clinical interpretability of the MAFM-Xception model for DR automatic grading were evaluated through visual analysis using gradient-weighted class activation mapping(Grad-CAM).Results The MAFM-Xception model achieved excellent classification performance on the APTOS-2019 blindness detection dataset,with an accuracy of 96.88％,a recall of 92.62％,a precision of 92.62％and an F1 score of 93.54％,which had the accuracy increased by 6.38 percentage points than the baseline Xception model and gained advantages in all the indexes over the mainsteam models such as ResNet50 and Inception-V3(all P＜0.001).Grad-CAM visualization results indicated the model accurately focused on key lesion areas in fundus images and demonstrated high interpretability.Conclusion The MAFM-Xception model exhibits relatively stable classification performance and is of practical value for DR automatic grading.[Chinese Medical Equipment Journal,2026,47(6):1-10]