SpurBreast
收藏资源简介:
SpurBreast是一个精心策划的乳腺MRI数据集,旨在研究现实世界中虚假相关性的影响。该数据集由真实世界的患者数据组成,故意包含了虚假相关性,以便评估其对模型性能的影响。通过分析超过100个涉及患者、设备和成像协议的特征,我们确定了两个主要的虚假信号:磁场强度(影响整个图像的全局特征)和图像方向(影响空间对齐的局部特征)。通过受控的数据集分割,我们证明DNN可以利用这些非临床信号,在验证集上实现高准确率,但在无偏见的测试数据上却无法泛化。
SpurBreast is a curated breast MRI dataset designed to investigate the impact of spurious correlations in real-world scenarios. This dataset comprises real-world patient data, and intentionally incorporates spurious correlations to evaluate their effects on model performance. By analyzing over 100 features related to patients, imaging devices and imaging protocols, we identified two primary spurious signals: magnetic field strength (a global feature affecting the entire image) and image orientation (a local feature affecting spatial alignment). Through controlled dataset splitting, we demonstrate that DNNs can leverage these non-clinical signals to achieve high accuracy on validation sets, but fail to generalize to unbiased test data.

- 1通过比利时根特大学 · 2025年



