1 from the virtual monochromatic X-ray image of the second energy E2, it is necessary to prepare in advance a trained model that can perform this inference. Therefore, first, a training phase for generating this trained model is described below. Furthermore, after describing the training phase, the flow of acquiring a high-quality first energy virtual monochromatic X-ray image according to the above principle is described.

In the training phase, first, original images are prepared which form a basis for generating training data. In the example embodiment, virtual monochromatic X-ray images of different energies Ei (i=1 to n) acquired by DECT are prepared as the original images.

In the present embodiment, we will consider two energies, the first energy E1 (i=1) and the second energy E2 (i=2), as the energy Ei of the virtual monochromatic X-ray image. The first energy E1 is 50 (keV) and the second energy E2 is 70 (keV). Therefore, a set of virtual monochromatic X-ray images V of the first energy E1 (=50 (keV)) and a set of virtual monochromatic X-ray images W of the second energy E2 (=70 (keV)) are prepared as the original images. Virtual monochromatic X-ray images of three or more energies E1 to En (n≥3) may also be prepared as the original images.

Next, pre-processing 68 is executed on these sets of virtual monochromatic X-ray images V and W, as illustrated in FIG. 4.

The pre-processing 68 includes, for example, image cropping, standardization, normalization, image inversion, image rotation, a magnification percentage change, and an image quality change. By pre-processing sets V and W of the original images, sets VA and WA of pre-processed virtual monochromatic X-ray images can be obtained. The preprocessed sets of virtual monochromatic X-ray images VA and WA are used as training data 60 to create the trained model.

Next, the training data 60 is used to train the neural network 70. The neural network 70 can use, for example, a convolutional neural network. In the present embodiment, the neural network 70 is trained to output a virtual monochromatic X-ray image with an energy of 50 (keV) based on a virtual monochromatic X-ray image with an energy of 70 (keV), and a trained model 71 is created.

After the trained model 71 is created, an evaluation of the trained model 71 is performed. The evaluation may, for example, use a confusion matrix. Accuracy, for example, may be used as an indicator index for the evaluation.

If the evaluation is favorable, the trained model 71 described above is used as the model for executing the inference (step ST3) described below. If the evaluation is poor, the training phase is performed again, for example, by adding more training data. This trained model 71 is stored in the storage device 218 of the CT system 100. The trained model 71 may be stored on an external storage device accessible by the CT system 100.

In the present embodiment, the trained model 71 is used to generate high-quality virtual monochromatic X-ray images. The flow for generating high-quality virtual monochromatic X-ray images using the trained model 71 is described below, with reference to the flow on the right side of FIG. 4. In step ST1, the imaging subject is scanned using DECT technology. Scans using DECT technology, can be, for example, performed using the Rapid kV switching method.

In step ST2, virtual monochromatic X-ray images of multiple energy Ei (i=1 to n) are reconstructed based on the data acquired in step ST1. The data obtained from the scan is collected by the DAS 214 (see FIG. 2), and the collected data is transmitted to the computer 216 or image reconstruction unit 230. On the computer 216 or image reconstruction unit 230, the processor reconstructs virtual monochromatic X-ray images of different energies based on the data obtained from the scan. In Embodiment 1, a series A of virtual monochromatic X-ray images a1 to az of the first energy E1 and a series B of virtual monochromatic X-ray images b1 to bz of the second energy E2 are reconstructed. Here, the first energy E1 is 50 (keV) and the second energy E2 is 70 (keV), but other energy values may be used. After reconstructing these virtual monochromatic X-ray image series A and B, the process proceeds to step ST3.

In step ST3, the computer 216 performs an inference phase to infer a series C of virtual monochromatic X-ray images c1 to cz of the first energy E1 (=50 keV) based on a series B of virtual monochromatic X-ray images b1 to bz of the second energy E2 (=70 keV). In the inference phase, the computer 216 pre-processes each virtual monochromatic X-ray image b1 to bz of the second energy E2 (=70 keV) to generate input images that are input to the trained model 71.

Next, the computer 216 inputs the input images to the trained model 71 and uses the trained model 71 to infer a series C of virtual monochromatic X-ray images c1 to cz of the first energy E1 (=50 keV). After inferring the image series C, the process proceeds to step ST4.

In step ST4, the computer 216 generates a corrected series D of virtual monochromatic X-ray images d1 to dz of the first energy E1 (=50 keV) based on the reconstructed series A of virtual monochromatic X-ray images a1 to az of the first energy E1 (=50 kev) and the inferred series C virtual monochromatic X-ray images c1 to cz of the first energy E1 (=50 keV).

This generation method is explained below with reference to FIG. 5. FIG. 5 is an explanatory diagram of a method of generating a corrected series D of virtual monochromatic X-ray images d1 to dz of the first energy E1 (=50 keV) in step ST4. In step ST41, the computer 216 extracts a single virtual monochromatic X-ray image aj from the reconstructed series A of virtual monochromatic X-ray images a1 to az of the first energy E1 (=50 kev). Here, assume that j=1, or in other words, the computer retrieves the virtual monochromatic X-ray image a1. After retrieving the virtual monochromatic X-ray image a1, processing proceeds to step ST42. In step ST42, the computer 216 identifies likely soft tissue areas in the virtual monochromatic X-ray image a1.

FIG. 6 is an explanatory diagram of the method for identifying tissue areas that are likely to be soft in step ST42. FIG. 6 illustrates two threshold values, TH1 and TH2, which are the criteria for determining whether a pixel CT value is a soft tissue CT value. The threshold value TH1 represents the lower limit of CT values for soft tissue, and the threshold value TH2 represents the upper limit of CT values for soft tissue. Therefore, the range G1 of TH1≤v≤TH2 represents the range where soft tissue CT values are likely to be included, and the range G2 of v<TH1 or TH2<v represents the range where soft tissue CT values are less likely to be included. Soft tissue CT values are generally included within the range of −150 (HU) to 150 (HU). Thus, TH1 and TH2 can be determined, for example, TH1=−150 (HU) and TH2=150 (HU).

The computer 216 compares the CT value v of each pixel of the virtual monochromatic X-ray image a1 with the threshold values TH1 and TH2 to determine, for each pixel, whether the CT value v is included in the range G1 of TH1≤v≤TH2 or the range G2 of v<TH1 or TH2<v.

For the sake of explanation, let us assume that the CT value v of pixels P11 to P1a in region R1 of the virtual monochromatic X-ray image a1 is TH1≤v≤TH2 and the CT value v of pixels P21 to P2b in region R2 is v<TH1 or TH2<v.

Therefore, the computer 216 identifies the region R1 including pixels P11 to Pla with CT value v in the range G1 of TH1≤v≤TH2 from the virtual monochromatic X-ray image a1 as the first region with a high possibility of soft tissue. On the other hand, the computer 216 identifies the region R2 including pixels P21 to P2b with CT values in the range G2 with v<TH1 or TH2<v from the virtual monochromatic X-ray image a1 as the second region with a low possibility of soft tissue.

Therefore, by executing the process of step ST42, pixels P11 to P1a that are more likely to belong to soft tissue and pixels P21 to P2b that are less likely to belong to soft tissue can be identified in the virtual monochromatic X-ray image a1.

Next, in step ST43, the computer 216 retrieves the virtual monochromatic X-ray image c1 from the inferred series C virtual monochromatic X-ray images c1 to cz of the first energy E1 (=50 keV), corresponding to the same cross-sectional image as the virtual monochromatic X-ray image a1. After retrieving the virtual monochromatic X-ray image c 1, processing proceeds to step ST44.

In step ST44, the computer 216 identifies areas that are likely soft tissue in the virtual monochromatic X-ray image c1. Specifically, the computer 216 identifies, from the virtual monochromatic X-ray image c1, the region R3 corresponding to the region R1 of the virtual monochromatic X-ray image a1 identified in step ST42 as the third region having a high possibility of being soft tissue. Furthermore, the computer 216 identifies, from the virtual monochromatic X-ray image c1, the region R4 corresponding to the region R2 in the virtual monochromatic X-ray image a1 identified in step ST42 as a fourth region having a low likelihood of being soft tissue.

Therefore, by executing the process of step ST44, pixels P31 to P3a that are more likely to belong to soft tissue and pixels P41 to P4b that are less likely to belong to soft tissue can be identified in the virtual monochromatic X-ray image c1.

The computer 216 then proceeds to step ST45. In step ST45, a corrected virtual monochromatic X-ray image of the first energy d1 is generated based on the reconstructed first energy E1 (=50 keV) virtual monochromatic X-ray image a1 and the inferred first energy E1 (=50 keV) virtual monochromatic X-ray image c1. An example of the flow of step ST45 is described below.

FIG. 7 illustrates an example of step ST45. Step ST45 includes steps ST451, ST452, and ST453. Each step is described in turn below. In step ST451, the weighting coefficients of the CT value v of each pixel in the virtual monochromatic X-ray images a1 and c1 is determined.

FIG. 8 is an explanatory diagram of how the weighting coefficients are determined. First, the weighting coefficient W1 of the CT value v of each pixel in region R1 of the reconstructed first energy E1 (=50 keV) virtual monochromatic X-ray image a1 and the weighting coefficient W3 of the CT value v of each pixel in region R3 of the inferred first energy E1 (=50 keV) virtual monochromatic X-ray image c1 are explained.

Pixels P11 to P1a in the region R1 of the reconstructed virtual monochromatic X-ray image a1 of the first energy E1 (=50 keV) are likely to correspond to soft tissue pixels since the CT value v is contained in TH1≤v≤TH2. In the low-energy virtual monochromatic X-ray image a1, the image quality of soft tissue tends to deteriorate easily, so region R1 is likely to be depicted with low image quality.

On the other hand, the inferred virtual monochromatic X-ray image c1 of the first energy E1 (=50 keV) inherits the favorable image quality of the virtual monochromatic X-ray image b1 (see FIG. 4) of the second energy E2 (=70 keV). Therefore, the image quality of region R3 of the inferred virtual monochromatic X-ray image c1 of the first energy E1 (=50 keV) is considered to be better than the image quality of region R1 of the reconstructed virtual monochromatic X-ray image a1 of the first energy E1 (=50 keV).

Therefore, to improve the image quality of soft tissue regions, pixels P11 to P1a in the region R1 of the reconstructed virtual monochromatic X-ray image a1 of the first energy E1 (=50 keV) are assigned a weighting coefficient W1 to lower the weighting of the CT values, considering that they are more likely to have low image quality. On the other hand, considering that pixels P31 to P3a in region R3 of the inferred virtual monochromatic X-ray image c1 of the first energy E1 (=50 keV) are better than the image quality of region R1 of the virtual monochromatic X-ray image a1, a weighting coefficient W3 is assigned to increase the weighting of the CT values.

Therefore, the weighting coefficient W3 for pixels P31 to P3a in region R3 of the inferred virtual monochromatic X-ray image c1 is assigned a value greater than the weighting coefficient W1 for pixels P11 to P1a in region R1 of the reconstructed virtual monochromatic X-ray image a1. Specifically, the values of W1 and W3 are set so that W3>W1 under the condition that W1+W3=1. For example, W1=0.3 and W3=0.7. In Embodiment 1, W1+W3=1, but the value of W1+W3 can be set to a value deviating from “1”.

Next, the weighting coefficient W2 of the CT value v of each pixel in region R2 of the reconstructed first energy E1 (=50 keV) virtual monochromatic X-ray image a1 and the weighting coefficient W4 of the CT value v of each pixel in region R4 of the inferred first energy E1 (=50 keV) virtual monochromatic X-ray image c1 are explained.

Since the CT value v of region R2 in the virtual monochromatic X-ray image a1 is included in the range of v<TH1 or TH2<v, air, gas, bone, and the like are considered to be depicted in region R