Author: Xiangli Cui, Man Hu, Wanli Huo, Da Yao, Jianguang Zhang, Yingying Zhang, Shanyang Zhao 👨🔬
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:
To study the fine-tuning strategy of pre-trained AI image generation model to adapt to the generation of small sample meningioma MRI images, explore its impact on observer performance, and determine the minimum data set for generating high-quality meningioma MRI images.
Methods:
To explore the feasibility of fine-tuning AI generation model for rare diseases with small data sets. The model has a CLIP text encoder and a noise prediction network ( U-Net ). By combining MSE loss and Adam optimizer, the parameters of convolution layer, normalization layer, skip connection layer and attention layer of U-net are adjusted to adapt to the radiation image domain, and the performance of fine-tuning strategy is evaluated by the coordination of multiple indicators. By measuring the signal-to-noise ratio ( SNR ) of human observers in the binomial forced-choice ( 2AFC ) test, the inseparability with real images is quantified, and the diversity of image quality is further quantified by t-SNE and MSSSIM.
Results:
The research on the relationship between the index and the number of samples shows that under the current situation of about 50 small sample sets, the meningioma images generated by the model adjusted by our method have higher quality and effectively reflect the different morphological characteristics of meningiomas. For simple tasks ( high-precision raw data conditions ), performance can still be retained when the number of current samples is one-tenth.
Conclusion:
This study demonstrates the potential in the field of MRI image generation for rare diseases. High-quality meningioma images can be generated using multiple sets of different numbers of far less than the traditional number of existing clinical real data, and the observer 's performance in all imaging tasks is unaffected. Future research can use this method in other medical imaging tasks.