Multi-level perturbations in image and feature spaces for semi-supervised medical image segmentation

Consistency regularization has emerged as a vital training strategy for semi-supervised learning. It is very important for medical image segmentation due to rare labeled data. To greatly enhance consistency regularization, we propose a novel Semi-supervised Learning framework with Multi-level Perturbations (SLMP) in both image and feature spaces. In image space, we propose external perturbations with three levels to greatly increase data variations. A low-level perturbation uses traditional augmentation techniques for firstly expanding data. Then, a middle-level one adopts copying and pasting techniques to combine low-level augmented versions of labeled and unlabeled data for generating new images. Middle-level perturbed images contain novel contents, which are totally different from original ones. Finally, a high-level one generates images from middle-level augmented data. In feature space, we design an Indicative Fusion Block (IFB) to propose internal perturbations for randomly mixing the encoded features of middle and high-level augmented images. By utilizing multi-level perturbations, we design a student–teacher semi-supervised learning framework for effectively improving the model resilience to strong variances. Experimental results show that our model achieves the state-of-the-art performance across various evaluation metrics on 2D and 3D medical image datasets. Our model exhibits the powerful capability of feature learning, and significantly outperforms existing state-of-the-art methods. Intensive ablation studies prove that our contributions are effective and significant. The model code is available at https://github.com/CamillerFerros/SLMP.