摘要
影像检查（如胸部X线摄影）通常会产生一组常见的影像学表现，以及数量更为庞大的罕见表现。尽管经验丰富的放射科医生可以通过学习少数具有代表性的罕见病例，掌握其视觉特征，但让机器从这种“长尾分布”中学习则困难得多，因为标准方法极易偏向于出现频率较高的类别。本文针对胸部X线图像中胸腔疾病领域的长尾学习问题，开展了一项全面的基准研究。我们聚焦于自然分布的胸部X线数据，旨在优化不仅在常见“头部”类别，而且在罕见但至关重要的“尾部”类别上的分类准确率。为实现这一目标，我们提出一个具有挑战性的新型长尾胸部X线图像基准数据集，以推动面向医学图像分类的长尾学习方法研究。该基准包含两个用于19类和20类胸腔疾病分类的胸部X线数据集，其中类别样本数量最多可达53,000张，最少仅有7张标注训练图像。我们在该新基准上评估了多种标准方法与当前最先进的长尾学习方法，深入分析了各类方法在长尾医学图像分类任务中的有效性，并总结出对未来算法设计具有指导意义的洞见。相关数据集、训练模型及代码已开源，详见：https://github.com/VITA-Group/LongTailCXR。
代码仓库
vita-group/longtailcxr
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| long-tail-learning-on-mimic-cxr-lt | MixUp | Balanced Accuracy: 0.176 |
| long-tail-learning-on-mimic-cxr-lt | Reweighted Focal Loss | Balanced Accuracy: 0.239 |
| long-tail-learning-on-mimic-cxr-lt | Class-balanced LDAM-DRW | Balanced Accuracy: 0.267 |
| long-tail-learning-on-mimic-cxr-lt | Class-balanced Softmax | Balanced Accuracy: 0.227 |
| long-tail-learning-on-mimic-cxr-lt | Reweighted Softmax | Balanced Accuracy: 0.211 |
| long-tail-learning-on-mimic-cxr-lt | Softmax | Balanced Accuracy: 0.169 |
| long-tail-learning-on-mimic-cxr-lt | Decoupling (tau-norm) | Balanced Accuracy: 0.230 |
| long-tail-learning-on-mimic-cxr-lt | Class-balanced Focal Loss | Balanced Accuracy: 0.191 |
| long-tail-learning-on-mimic-cxr-lt | Class-balanced LDAM | Balanced Accuracy: 0.225 |
| long-tail-learning-on-mimic-cxr-lt | Reweighted LDAM | Balanced Accuracy: 0.243 |
| long-tail-learning-on-mimic-cxr-lt | Balanced-MixUp | Balanced Accuracy: 0.168 |
| long-tail-learning-on-mimic-cxr-lt | LDAM | Balanced Accuracy: 0.165 |
| long-tail-learning-on-mimic-cxr-lt | Focal Loss | Balanced Accuracy: 0.172 |
| long-tail-learning-on-mimic-cxr-lt | Reweighted LDAM-DRW | Balanced Accuracy: 0.275 |
| long-tail-learning-on-mimic-cxr-lt | Decoupling (cRT) | Balanced Accuracy: 0.296 |
| long-tail-learning-on-nih-cxr-lt | Reweighted Focal Loss | Balanced Accuracy: 0.197 |
| long-tail-learning-on-nih-cxr-lt | Focal Loss | Balanced Accuracy: 0.122 |
| long-tail-learning-on-nih-cxr-lt | Balanced-MixUp | Balanced Accuracy: 0.155 |
| long-tail-learning-on-nih-cxr-lt | Softmax | Balanced Accuracy: 0.115 |
| long-tail-learning-on-nih-cxr-lt | Reweighted LDAM | Balanced Accuracy: 0.279 |
| long-tail-learning-on-nih-cxr-lt | LDAM | Balanced Accuracy: 0.178 |
| long-tail-learning-on-nih-cxr-lt | Class-balanced LDAM | Balanced Accuracy: 0.235 |
| long-tail-learning-on-nih-cxr-lt | Decoupling (cRT) | Balanced Accuracy: 0.294 |
| long-tail-learning-on-nih-cxr-lt | Reweighted LDAM-DRW | Balanced Accuracy: 0.289 |
| long-tail-learning-on-nih-cxr-lt | Decoupling (tau-norm) | Balanced Accuracy: 0.214 |
| long-tail-learning-on-nih-cxr-lt | Class-Balanced Focal Loss | Balanced Accuracy: 0.232 |
| long-tail-learning-on-nih-cxr-lt | MixUp | Balanced Accuracy: 0.118 |
| long-tail-learning-on-nih-cxr-lt | Class-balanced LDAM-DRW | Balanced Accuracy: 0.281 |
| long-tail-learning-on-nih-cxr-lt | Reweighted Softmax | Balanced Accuracy: 0.260 |
| long-tail-learning-on-nih-cxr-lt | Class-Balanced Softmax | Balanced Accuracy: 0.269 |