Abstract: In clinical applications,CT image is a kind of image data that is relatively easy to obtain,but there is a large gap between them and the real human body color.The tomographic color image of the human body is the color response of the real human body,but it is a rare data.Combining the two,so that each case can get its own color CT data,which will have a effect on the doctor’s surgery and the patient’s understanding to the disease.Therefore,this paper proposes a medical image colorization framework based on a color flow model.It first inputs the CT and human color data into the color flow model and extracts the content and color features.Then,the color and texture transfer work is performed at the feature level.Finally,the processed feature information is re-input into the reversible color flow model for image reconstruction.After each flow module,we add a texture constraint loss to make the shaded image more textured.At the same time,we add edge constraints to ensure that the characteristics of small blood vessels and other tissues on the medical image are not lost.Qualitative and quantitative experiments prove that our method is more robust than other colorization methods,and the experimental results are more realistic.And we conduct extensive experiments on different data domains,proving that our method is not affected by domain shift and can obtain stable experimental results.At the same time,the proposed method can display a clear organizational structure without adjusting the window width/level.

Key words: Flow module, Colorization, Texture constraint, Stability, Edge constraint