Voxel-Wise Medical Imaging Transformation and Adaption Based on CycleGAN and Score-Based Diffusion

Area of Science:

• Medical imaging
• Artificial intelligence
• Computer vision

Background:

• Supervised learning in medical imaging degrades with non-independent and identically distributed (i.i.d.) data.
• Distribution shifts between training and testing datasets from different CT scanner manufacturers pose a significant challenge.

Purpose of the Study:

• To develop a method for harmonizing medical imaging data from diverse sources.
• To improve the fidelity of generated medical images by mitigating artifacts.
• To enhance the performance of supervised learning models on heterogeneous datasets.

Main Methods:

• Utilized CycleGAN (Generative Adversarial Networks) for initial domain adaptation of Computer Tomography (CT) data.
• Employed a score-based generative model to refine images, removing artifacts and boundary marks.
• Developed a hybrid generative model approach for cross-manufacturer CT data transformation.

Main Results:

• Successfully reduced distribution shifts between CT datasets from different terminals/manufacturers.
• Mitigated radiology artifacts common in GAN-generated images.
• Achieved higher fidelity image transformation without sacrificing significant features.

Conclusions:

• The combined generative model approach effectively addresses data heterogeneity in medical imaging.
• This method shows promise for improving the robustness of supervised learning on diverse datasets.
• Future work will involve evaluating a wider range of supervised methods on both original and generated datasets.