📄 中文摘要
Oculomics即通过视网膜图像预测心血管疾病和痴呆等全身性疾病，在RETFound等基于Transformer的预训练模型推动下取得了快速进展。图像级别的混合样本数据增强技术，如CutMix和MixUp，常用于Transformer模型的训练。然而，这些技术会扰乱患者特定的属性，例如医疗合并症和临床事实。当使用这些增强技术时，模型可能会学习到与患者特定属性不一致的混合标签，从而导致训练效率低下，甚至损害模型的性能。为解决这一问题，引入oculomix，一种新的层级采样策略。该方法在混合样本增强过程中，通过对患者层面的特征进行分层处理，确保生成混合图像的标签能够更准确地反映其潜在的临床意义。
📄 English Summary
oculomix: Hierarchical Sampling for Retinal-Based Systemic Disease Prediction
Oculomics, the concept of predicting systemic diseases such as cardiovascular disease and dementia through retinal imaging, has advanced rapidly, largely driven by the data efficiency of transformer-based foundation models like RETFound. Image-level mixed sample data augmentations, such as CutMix and MixUp, are frequently employed for training these transformer models. However, these techniques inherently perturb patient-specific attributes, including medical comorbidities and clinical facts. When these augmentation methods are applied, models may learn mixed labels that are inconsistent with the underlying patient-specific characteristics, potentially leading to inefficient training or even degradation of model performance. To address this critical limitation, oculomix is introduced as a novel hierarchical sampling strategy. This method ensures that the labels of generated mixed images more accurately reflect their underlying clinical significance by performing hierarchical processing of patient-level features during the mixed sample augmentation process. Specifically, oculomix not only considers pixel-level mixing when creating new training samples but, more importantly, strategically combines metadata at the patient or sample level, thereby avoiding the semantic inconsistencies that traditional mixing methods might introduce. This approach enables the model to better understand the complex associations between retinal images and systemic diseases while preserving the clinical fidelity of the dataset. By adopting this strategy, oculomix aims to enhance the accuracy and generalization capabilities of models in predicting systemic diseases, allowing them to extract valuable biomedical information from retinal images more effectively and thereby fostering advancements and clinical applications in the field of Oculomics.