Image processing apparatus, image processing method, and X-ray CT apparatus
Noise is reduced for a medical image for which noise cannot be quantified by a general-purpose image quality evaluation index. An image processor has a preprocessor that generates input images including an original image and one or more images with reduced noise compared with the original image; and a noise reduction processor outputs an image, which is obtained by reducing noise from the original image based on the input images, by applying a learned network. The learned network used in the noise reduction processor is constructed by performing deep learning using a plurality of learning sets in which one or more of a medical image including noise, a noise-reduced image obtained by performing noise reduction processing on the medical image, and an intermediate image obtained during the noise reduction processing are input images and a correct image is obtained based on the input images an output image.
1. An image processing apparatus, comprising:
a preprocessing unit which includes a noise reduction processing unit; and
a learning application unit that applies a learned network to perform noise reduction processing of an original image,
wherein the preprocessing unit outputs the original image to the learning application unit as a first input image among a plurality of input images input to the learning application unit,
wherein the noise reduction processing unit generates a plurality of second input images among the plurality of input images to be input to the learning application unit by performing preprocessing noise reduction processing on the original image to obtain the plurality of second input images,
wherein the learning application unit performs the noise reduction processing of the original image based on the plurality of input images including the first input image and the plurality of second input images,
wherein the learned network is constructed by performing deep learning using a plurality of learning sets of the plurality of input images which include two or more of the original image, a metal image having a value only in a metal region, a linear interpolation image obtained by linearly interpolating components affected by metal in a projection space, a beam hardening correction image in which noise caused by metal has been reduced by beam hardening correction processing, a high frequency image in which high frequency components of the original image are emphasized, an edge image showing a magnitude of a difference in pixel value between adjacent pixels in the original image, and a region divided image obtained by dividing a region according to a difference in X-ray attenuation coefficient of an imaging target, and
wherein the learned network generates a correct image, which is obtained based on the plurality of input images and in which noise caused by metal has been reduced, as an output image.
2. The image processing apparatus according to claim 1 ,
wherein the correct image is obtained by performing, on the input images, any processing of a weighted sum in a real space, a weighted sum in a frequency space, and a combination of the weighted sums in the real space and the frequency space.
3. The image processing apparatus according to claim 1 ,
wherein the correct image is obtained by performing known noise reduction processing on the input images in a state in which a value of a predetermined parameter used for processing has been specified by a user.
4. An image processing method, comprising:
a preprocessing step for outputting an original image as a first input image among a plurality of input images to be input to a learning application unit;
a noise reduction processing step for generating a plurality of second input images among the plurality of input images to be input to a learning application unit by performing preprocessing noise reduction on the original image to obtain the plurality of second input images;
applying, by the learning application unit, a learned network to perform noise reduction processing of the original image based on the plurality of input images,
wherein the learned network is constructed by performing deep learning using a plurality of learning sets in which two or more of the original image, a metal image having a value only in a metal region, a linear interpolation image obtained by linearly interpolating components affected by metal in a projection space, a beam hardening correction image in which noise caused by metal has been reduced by beam hardening correction processing, a high frequency image in which high frequency components of the original image are emphasized, an edge image showing a magnitude of a difference in pixel value between adjacent pixels in the original image, and a region divided image obtained by dividing a region according to a difference in X-ray attenuation coefficient of an imaging target, and
generating a correct image using the learned network, which is obtained based on the plurality of input images and in which noise caused by metal has been reduced, as an output image.
5. An X-ray CT apparatus, comprising:
an imaging unit that emits X-rays to a subject and detects X-rays transmitted through the subject to generate an image; and
the image processing apparatus according to claim 1 ,
wherein the image processing apparatus performs noise reduction processing using the image generated by the imaging unit as the original image.