CFAB: An Online Data Augmentation to Alleviate the Spuriousness of Classification on Medical Ultrasound Images

Convolutional neural networks (CNNs) may learn spurious correlations between bias features (e.g., background) and labels in image classification. The spuriousness in CNNs usually occurs in building connections between the background of images and labels. Such spurious correlation limits the generalizability of CNNs in classification tasks. Changing backgrounds and foregrounds of original samples can reduce the spuriousness in natural image classification. However, generating annotations for foreground on medical image datasets is time-consuming and labor-intensive. To solve this problem, we propose an online data augmentation method named Combining Foreground And Background (CFAB), which makes CNNs focus on key causal features without foreground annotations and breaks the correlation between backgrounds and labels by changing different backgrounds for one sample. Furthermore, we propose a framework for collaborative augmenting samples using CFAB and training CNNs. Comprehensive experiments indicate that the proposed method weakens the spuriousness, improves the generalizability of the model, and achieves state-of-the-art results in medical ultrasound dataset classification.