Data augmentation strategies for semi-supervised medical image segmentation

Abstract

Exploiting unlabeled and labeled data augmentations has become considerably important for semi-supervised medical image segmentation tasks. However, existing data augmentation methods, such as Cut-mix and generative models, typically dependent on consistency regularization or ignore data correlation between slices. To address cognitive biases problems, we propose two novel data augmentation strategies and a Dual Attention-guided Consistency network (DACNet) to improve semi-supervised medical image segmentation performance significantly. For labeled data augmentation, we randomly crop and stitch annotated data rather than unlabeled data to create mixed annotated data, which breaks the anatomical structures and introduces voxel-level uncertainty in limited annotated data. For unlabeled data augmentation, we combine the diffusion model with the Laplacian pyramid fusion strategy to generate unlabeled data with higher slice correlation. To enhance the decoders to learn different semantic but discriminative features, we propose the DACNet to achieve structural differentiation by introducing spatial and channel attention into the decoders. Extensive experiments are conducted to show the effectiveness and generalization of our approach. Specifically, our proposed labeled and unlabeled data augmentation strategies improved accuracy by 0.3% to 16.49% and 0.22% to 1.72%, respectively, when compared with various state-of-the-art semi-supervised methods. Furthermore, our DACNet outperforms existing methods on three medical datasets (91.72% dice score with 20% labeled data on the LA dataset). Source code will be publicly available at https://github.com/Oubit1/DACNet.