Author: Xiangli Cui, Zilei Fu, Man Hu, Wanli Huo, Xiaoqing Wu, Jianguang Zhang, Yingying Zhang 👨🔬
Affiliation: Institute of Health and Medical Technology, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Hefei, the Zhejiang-New Zealand Joint Vision-Based Intelligent Metrology Laboratory, College of Information Engineering, China Jiliang University, Departments of Radiation Oncology, Zibo Wanjie Cancer Hospital, Department of Oncology, Xiangya Hospital, Central South University, Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University, Shandong Academy of Medical Sciences 🌍
Purpose:
Using Stable Diffusion to generate images of the knee in different disease states can enrich the medical imaging database and inject new vitality into the field of medical imaging analysis. These images not only have high authenticity and detail representation, but also can cover various situations from mild lesions to serious diseases, meeting the diverse data needs of medical imaging research.
Methods:
Build a knee model and study the accuracy of knee generation. The method parameters include the number of sampling steps (usually set at 20 to 50 steps, which affects the quality and details of the image, with higher steps having more details but longer generation time), sampler (commonly Euler a, DDIM, etc., with different effects), and resolution (set according to the desired image size, with higher resolution details but higher hardware requirements). By modifying parameters to generate images, the results of the images were evaluated.
Results:
Stable Diffusion technology enriches medical imaging databases, enhancing image authenticity and detail expression. The evaluation shows that the average SSIM of the image has increased by 0.25 and the PSNR has increased by 5dB, indicating optimization of the image structure and signal quality. Training a disease diagnosis model with this database resulted in a 10% increase in accuracy and an 8% increase in recall, demonstrating its outstanding performance in medical image enhancement and assisting diagnostic accuracy.
Conclusion:
This work demonstrates significant effectiveness in using Stable Diffusion to generate images of the knee under different disease states. By setting parameters such as sampling steps, sampler, and resolution reasonably, knee images can be generated more accurately. With these precise images, knee image segmentation medical image data can be increased, solving the problem of insufficient medical image data faced by deep learning based knee image segmentation.