Synthesizing Images With Annotations for Medical Image Segmentation Using Diffusion Probabilistic Model

Abstract

To alleviate the burden of manual annotation, there are numerous excellent segmentation models for images segmentation being developed. However, the performance of these data-driven segmentation models is frequently constrained by the availability of samples sizes of pair medical images and segmentation annotations. Therefore, to address this challenge, this study introduces the medical image segmentation augmentation diffusion model (MEDSAD). MEDSAD solves the problem of annotation scarcity by utilizing a given simple annotation to generate paired medical images. To improve stability, we used the traditional diffusion model for this study. To exert better control over the texture synthesis in the medical images generated by MEDSAD, the texture style injection (TSI) mechanism is introduced. Additionally, we propose the feature frequency domain attention (FFDA) module to mitigate the adverse effects of high-frequency noise during generation. The efficacy of MEDSAD is substantiated through the validation of three distinct medical segmentation tasks encompassing magnetic resonance (MR) and ultrasound (US) imaging modalities, focusing on the segmentation of breast tumors, brain tumors, and nerve structures. The findings demonstrate the MEDSAD model's proficiency in synthesizing medical image pairs based on provided annotations, thereby facilitating a notable augmentation in performance for subsequent segmentation tasks. Moreover, the improvement in performance becomes greater as the quantity of synthetic available data samples increases. This underscores the robust generalization capability and efficacy intrinsic to the MEDSAD model, potentially offering avenues for future explorations in data-driven model training and research.