预处理方法对糖尿病视网膜病变分级中单源域泛化影响的实证研究

An Empirical Study on the Impact of Preprocessing Methods on Single-Source Domain Generalization in Diabetic Retinopathy Grading

摘要: 针对糖尿病视网膜病变(DR)自动诊断中跨数据集泛化能力不足的问题，文章开展了一项关于预处理方法对单源域泛化性能影响的实证研究。以EfficientNet为统一骨干网络，在APTOS数据集上训练，在DDR、IDRiD和ISBI三个独立数据集上进行零样本泛化测试，系统对比了五种预处理策略：无预处理、Circle Crop + CLAHE、仅MixStyle、仅DRGen、CLAHE + MixStyle。实验结果表明，不同预处理方法对源域精度与目标域泛化能力的影响存在显著差异，其中Circle Crop + CLAHE组合在保持源域性能的同时，在三个目标域上取得最优的平均泛化性能，而MixStyle和DRGen等风格扰动方法也展现出良好的跨域适应能力。文章首次在DR分级跨域场景下对主流预处理方法进行系统性横向对比，揭示了预处理选择对单源域泛化的关键影响，为多中心DR筛查系统的构建提供了可量化的预处理参考依据。

Abstract: This paper presents an empirical study on the impact of preprocessing methods on single-source domain generalization in response to the issue of insufficient cross-dataset generalization capability in automatic diabetic retinopathy (DR) diagnosis. Using EfficientNet as the unified backbone network, models were trained on the APTOS dataset and subjected to zero-shot generalization testing on three independent datasets—DDR, IDRiD, and ISBI—to systematically compare five preprocessing strategies: no preprocessing, Circle Crop with CLAHE, MixStyle only, DRGen only, and CLAHE with MixStyle. Experimental results reveal significant variations in the effects of different preprocessing methods on source domain accuracy and target domain generalization. Among these, the combination of Circle Crop + CLAHE achieves the best average generalization performance across the three target domains while preserving source domain performance. Additionally, style perturbation methods such as MixStyle and DRGen demonstrate strong cross-domain adaptation capability. This study presents the first systematic horizontal comparison of mainstream preprocessing methods in the context of cross-domain DR grading, highlighting the critical role of preprocessing choices in single-source domain generalization and providing a quantifiable reference for the development of multi-center DR screening systems.