Double-energy-spectrum CT projection domain base material decomposition method and device based on deep learning

Abstract

The invention discloses a multi-energy-spectrum CT projection domain base material decomposition method and device based on deep learning, and the method comprises the steps: employing multi-energy-spectrum CT to collect the multi-energy-spectrum multi-color projection of a known calibration model body in a network training process; using multicolor projection under an energy spectrum to directlyreconstruct a mold body CT image, segmenting the mold body CT image into a plurality of base material images, and respectively solving line integrals of the base material images along each ray; designing a deep neural network for multi-color projection decomposition, taking multi-energy-spectrum multi-color projection of the calibration model body as network input, and taking line integral of a base material image of the calibration model body as an output label to complete training; in the network application process, inputting multi-energy-spectrum multi-color projection of a measured objectinto a neural network, and decomposing line integration of a multi-base material image; and reconstructing a multi-base material density image of the measured object through the line integrals. The method is good in anti-noise performance, has no strict requirement on the correlation between the calibration mold body and the morphological structure of the measured object, and does not need to measure energy spectrum information in advance.

Description

technical field [0001] The present invention relates to the technical field of X-ray multi-energy spectrum CT imaging, in particular to a deep learning-based multi-energy spectrum CT-based material image decomposition method. Background technique [0002] X-ray CT imaging technology is a technology that uses the principle of interaction between X-rays and matter to image the internal information of the scanned object. The X-rays emitted by conventional CT ray sources have different wavelengths and frequencies, and obey a certain continuous polychromatic energy spectrum distribution like visible light. Conventional CT collects the polychromatic projection of the X-ray energy spectrum passing through the measured object, rather than the ideal single-energy X-ray projection. Images reconstructed using traditional single-energy CT reconstruction algorithms not only have significant hardening artifacts, but also different substances may have the same or similar CT values, making...

AI Technical Summary

Problems solved by technology

When the convolutional neural network processes each pixel of the image, it uses a large amount of neighborhood information, which has the advantage of good denoising performance, but at the same time it causes a disadvantage, that is, the difference between the samples for training the network and the image of the object to be processed by the network must be very close in structure

Since it is difficult to obtain a large amount of information of the image to be processed in advance to train the convolutional neural network in many cases, the scope of application of this type of method is severely limited.

Images