In medical image segmentation tasks, the domain shift problem affects the performance of trained segmentation models in the unseen domain, therefore, improving model generalization is crucial for the practical application of intelligent models for medical images. Representation learning is currently one of the dominant methods for solving domain generalization problems, mostly using image-level loss and consistency loss to supervise image generation. However, it is not sensitive enough to the deviation of small morphological features of medical images, which can lead to unclear edges of the generated images and affect the subsequent learning of the model. In order to improve the generalization of the model, this paper proposes a semi-supervised Feature-level Loss and Learnable Noise Domain Generalization (FLLN-DG) method for medical image segmentation. Firstly the introduction of feature level loss improves the problem of unclear boundaries of the generated images. Secondly the introduction of learnable noise components further increases the data diversity and improves the model generalization. Compared with the baseline model, FLLN-DG improves the performance in the unseen domain by 2% to 4%, which demonstrates the effectiveness of feature-level loss and learnable noise components. FLLN-DG also has the best generalization and segmentation results compared to typical generalization models such as nnUNet, SDNet+AUG, LDDG, SAML, and Meta.

[2023-11-16] Accepted Paper /doc/2023.08.0408/v1