CF-ResNet：融合注意力与多尺度特征的糖尿病视网膜病变分级诊断模型研究

CF-ResNet: Research on Grading Diagnosis Model of Diabetic Retinopathy Fused with Attention and Multi-Scale Features

摘要: 针对糖尿病视网膜病变(DR)自动诊断中存在的病灶特征识别不精准、数据分布不均衡及模型泛化能力不足等问题，我们提出一种基于ResNet50的改进模型CF-ResNet。以Kaggle公开眼底图像数据集为研究对象，通过多维度优化策略提升模型诊断性能：引入CBAM注意力机制强化对微小病灶的特征聚焦能力；采用Focal Loss损失函数缓解数据类别不平衡带来的训练偏差；结合多种针对性数据增强方法扩充有效样本并提升模型鲁棒性；新增多尺度特征融合模块(MSFM)适配不同尺寸病变的特征提取需求。实验结果表明，CF-ResNet模型在测试集上的准确率达90.3%、召回率为90.6%、特异性为92.3%、F1分数为90.4%，各项指标均优于原始ResNet50及主流对比模型。消融实验验证了各改进模块的有效性，模型在普通设备上单张图像推理耗时仅0.06秒，具备临床辅助诊断与大规模筛查的应用潜力。

Abstract: Aiming at the problems of inaccurate lesion feature recognition, unbalanced data distribution and insufficient model generalization ability in the automatic diagnosis of Diabetic Retinopathy (DR), an improved model CF-ResNet based on ResNet50 is proposed. Taking the public Kaggle fundus image dataset as the research object, the diagnostic performance of the model is improved through multi-dimensional optimization strategies: introducing the CBAM attention mechanism to strengthen the feature focusing ability on micro lesions; adopting the Focal Loss function to alleviate the training bias caused by unbalanced data categories; combining a variety of targeted data augmentation methods to expand effective samples and improve model robustness; adding a Multi-Scale Feature Fusion Module (MSFM) to adapt to the feature extraction of lesions of different sizes. Experimental results show that the CF-ResNet model achieves an accuracy of 90.3%, a recall rate of 90.6%, a specificity of 92.3%, and an F1 score of 90.4% on the test set, and all indicators are superior to the original ResNet50 and mainstream comparison models. Ablation experiments verify the effectiveness of each improved module, and the average inference time of the model for a single image on ordinary equipment is only 0.06 seconds, which has the potential for clinical auxiliary diagnosis and large-scale screening.