Four Public Datasets for Explainable Medical Image Classifications

Modern deep neural networks are overparameterized and thus require data augmentation techniques to prevent over-fitting and improve generalization ability. Generative adversarial networks (GANs) are famous for generating visually realistic images. However, the generated images lack diversity and have uncertain class labels. On the other hand, recent methods mix labels proportionally to the salient region. As the main diagnostic information is not always contained within the salient region, we argue that methods along this line can cause label mismatch issues in medical image classifications. Therefore, we propose VariMix, which exploits an absolute difference map (ADM) to address the label mismatching of mixed medical images. The VariMix generates ADM using the image-to-image (I2I) GAN across multiple classes and allows for bidirectional mixing operations between the training samples.  We collect four public medical image datasets for automatic medical image classifications: Breast Ultrasound datset, Chest X-Ray Images (CXR) dataset, Eye Disease Retinal Images (Retinal) dataset, Maternal-fetal ultrasound dataset. Extensive experiments prove the superiority of VariMix compared with the existing GAN-based and Mixupbased augmentation methods on four public datasets using Swin Transformer V2 and ConvNeXt architectures. Furthermore, by projecting the source image to the hyperplane of the support vector machine, the proposed I2I GAN can generate hyperplane difference maps (HDM) between the source image and the hyperplane image, demonstrating its ability to interpret medical image classifications. 

现代深度神经网络存在过参数化问题，因此需要借助数据增强技术来避免过拟合并提升泛化能力。生成式对抗网络（Generative Adversarial Networks, GANs）因能够生成视觉效果逼真的图像而广为人知，但这类网络生成的图像往往多样性不足，且类别标签存在不确定性。另一方面，现有部分方法会根据显著区域按比例混合标签。由于核心诊断信息并非始终集中于显著区域，我们认为这类方法会在医学图像分类任务中引发标签不匹配问题。为此，我们提出了VariMix方法，该方法通过绝对差异图（Absolute Difference Map, ADM）来解决混合医学图像的标签不匹配问题。VariMix借助跨多类别图像到图像（Image-to-Image, I2I）生成式对抗网络生成绝对差异图，并支持训练样本间的双向混合操作。我们收集了四个用于自动医学图像分类的公开医学图像数据集：乳腺超声数据集、胸部X线影像（Chest X-Ray Images, CXR）数据集、眼部疾病视网膜影像（Retinal）数据集以及母胎超声数据集。我们基于Swin Transformer V2与ConvNeXt两种架构，在四个公开数据集上开展了大量对比实验，结果证明VariMix的性能优于现有基于生成式对抗网络和Mixup的数据增强方法。此外，通过将源图像投影至支持向量机（Support Vector Machine, SVM）的超平面，我们所提出的图像到图像生成式对抗网络能够生成源图像与超平面图像间的超平面差异图（Hyperplane Difference Map, HDM），从而验证了其可解释医学图像分类的能力。