An unsupervised/semi-supervised CT image reconstruction depth network train method

Abstract

An unsupervised / semi-supervised CT image reconstruction depth network train method is provided. Firstly, the CT chord data under the CT scanning protocol and the imaging system parameters of the CT equipment are obtained. The data include low-dose CT chord data without labeling, and a small amount of low-dose CT chord data correspond to labeled CT chord data. The labeled CT chord data refers to the clear CT chord data information of known low dose data corresponding to high dose. The loss functions of unlabeled and labeled data are constructed respectively, and the total loss functions of unsupervised / semi-supervised network are obtained by weighted summation, and the denoising network is trained by this loss function. The method achieves the effect of a CT image denoising network with higher precision and higher speed by using non-annotated data only or only a small amount of annotated data. The invention aims at establishing a method of combining a chord data restoration model and adepth learning model, thereby realizing high-quality CT image reconstruction.

Description

technical field [0001] The invention belongs to the technical field of image processing, and relates to image processing of medical images, in particular to an unsupervised / semi-supervised CT image reconstruction deep network training method based on fusion of deep learning and error modeling framework. Background technique [0002] CT (Computed Tomography) is currently widely used in clinical medical imaging diagnosis. However, excessive doses of X-rays will cause potential harm to the human body, easily induce malignant tumors, and cause organ damage. Therefore, reducing the dose of X-rays as much as possible has become one of the key technologies in the field of medical CT imaging. However, this It often leads to severe degradation and serious noise in CT chord data. [0003] In order to solve the problem of noise contained in low-dose CT images, two common methods have been proposed, one is the method of modeling the noise and constructing the maximum a posteriori model...

AI Technical Summary

Problems solved by technology

[0004] The method based on the maximum a posteriori model has higher accuracy in actual use, but due to the existence of a large number of iterative processes in the algorithm, the calculation cost is high and the time-consuming is long (tens or even hundreds of times the denoising speed of deep learning) , At the same time, since this method can only process one CT image at a time, it lacks the extraction of common features, which also limits the best effect that this method can achieve

[0005] The method based on deep learning is fast and works well in the case of a large amount of accurate labeled data, but the acquisition cost of labeled data is high and the damage to the patient is great, so it is often difficult to obtain a large amount of labeled data in practical applications, and due to Annotated data is generally obtained from high-dose data simulation to obtain corresponding low-dose data, and the simulation process of low-dose CT chord diagram itself will also introduce a lot of noise, and the noise of the data will also have a great impact on the accuracy of the network

In general, the quantity and quality of labeled data will limit the performance of deep learning in practical applications.

Images