Medical image processing method for processing pediatric simple x-ray image using machine learning model and medical image processing apparatus therefor
A medical image processing method for processing a pediatric simple X-ray image using a machine learning model and a medical image processing apparatus therefor are provided. The apparatus applies a style conversion model to a CT image pair including a basic CT image and a suppression CT image where at least a portion of a bone of the basic CT image is suppressed to convert the CT image pair into a conversion image pair and trains a bone suppression model for a simple X-ray image based on training data including the conversion image pair.
1. A method performed by a processor of a medical image processing apparatus, the method comprising:
obtaining a basic computer tomography (CT) image and a suppression CT image where at least a portion of a bone of the basic CT image is suppressed;
applying a style conversion model to a CT image pair including the basic CT image and the suppression CT image to convert the CT image pair into a conversion image pair; and
training a bone suppression model for a simple X-ray image based on training data including the conversion image pair,
wherein the obtaining of the basic CT image and the suppression CT image includes:
projecting a basic stereoscopic CT image onto a reference plane to obtain the basic CT image being a planar CT image; and
projecting a suppression stereoscopic CT image onto the reference plane to obtain the suppression CT image being a planar CT image, the suppression stereoscopic CT image being obtained by changing a first voxel value corresponding to at least a portion of a bone of the basic stereoscopic CT image to a second voxel value corresponding to a non-bone, the at least a portion of the bone being a bone corresponding to a voxel having the first voxel value which belongs to a threshold range in the bone of the basic stereoscopic CT image, and the second voxel value being a value of a voxel corresponding to the non-bone within a threshold distance from a voxel having the first voxel value,
wherein the training of the bone suppression model includes:
applying the bone suppression model to a basic conversion image of the conversion image pair to obtain a training output image;
calculating an objective function value of the bone suppression model based on the training output image and a suppression conversion image of the conversion image pair; and
updating parameters of the bone suppression model using the objective function value,
wherein the calculating of the objective function value includes:
calculating the objective function value based on a pixel value difference between a first pixel of the training output image and a second pixel of the suppression conversion image, the second pixel corresponding to the first pixel, the objective function value being calculated as a large value as the pixel value difference increases, and
wherein the obtaining of the training output image, the calculating of the objective function value, and the updating of the parameters are repeated to train the bone suppression model, until the objective function value is less than a predetermined threshold.
2. The method of claim 1 , further comprising:
applying the trained bone suppression model to a simple X-ray input image to obtain a simple X-ray output image where at least a portion of a bone of the simple X-ray input image is suppressed.
3. The method of claim 2 , wherein the obtaining of the simple X-ray output image includes:
applying the bone suppression model to the simple X-ray input image for pediatrics to obtain the simple X-ray output image where the at least a portion of the bone of the simple X-ray input image is suppressed.
4. The method of claim 1 , further comprising:
obtaining the CT image pair for pediatrics, the CT image pair including the basic CT image and the suppression CT image for the pediatrics.
5. The method of claim 1 , wherein the converting into the conversion image pair includes:
applying the style conversion model to the CT image pair for pediatrics to convert the CT image pair for the pediatrics into the conversion image pair for the pediatrics.
6. The method of claim 1 , wherein the training of the bone suppression model includes:
using a basic conversion image of the conversion image pair of the training data as a training input image and using a suppression conversion image of the conversion image pair as a ground truth output image to train the bone suppression model by means of supervised learning.
7. A computer-readable storage medium storing a program combined with hardware to execute the method of claim 1 .
8. An apparatus, comprising:
a processor configured to obtain a basic CT image and a suppression CT image where at least a portion of a bone of the basic CT image is suppressed, apply a style conversion model to a CT image pair including the basic CT image and the suppression CT image to convert the CT image pair into a conversion image pair, and train a bone suppression model for a simple X-ray image based on training data including the conversion image pair,
wherein the processor is configured to:
project a basic stereoscopic CT image onto a reference plane to obtain the basic CT image being a planar CT image, when obtaining the basic CT image and the suppression CT image, and project a suppression stereoscopic CT image onto the reference plane to obtain the suppression CT image being a planar CT image, the suppression stereoscopic CT image being obtained by changing a first voxel value corresponding to at least a portion of a bone of the basic stereoscopic CT image to a second voxel value corresponding to a non-bone, the at least a portion of the bone being a bone corresponding to a voxel having the first voxel value which belongs to a threshold range in the bone of the basic stereoscopic CT image, and the second voxel value being a value of a voxel corresponding to the non-bone within a threshold distance from a voxel having the first voxel value;
apply the bone suppression model to a basic conversion image of the conversion image pair to obtain a training output image, when training the bone suppression model, calculate an objective function value of the bone suppression model based on the training output image and a suppression conversion image of the conversion image pair, and update parameters of the bone suppression model using the objective function value;
calculate the objective function value based on a pixel value difference between a first pixel of the training output image and a second pixel of the suppression conversion image, the second pixel corresponding to the first pixel, when calculating the objective function value, the objective function value being calculated as a large value as the pixel value difference increases; and
repeat the obtaining of the training output image, the calculating of the objective function value, and the updating of the parameters to train the bone suppression model, until the objective function value is less than a predetermined threshold.
9. The apparatus of claim 8 , wherein the processor applies the trained bone suppression model to a simple X-ray input image to obtain a simple X-ray output image where at least a portion of a bone of the simple X-ray input image is suppressed.
10. The apparatus of claim 9 , wherein the processor applies the bone suppression model to the simple X-ray input image for pediatrics to obtain the simple X-ray output image where the at least a portion of the bone of the simple X-ray input image is suppressed.
11. The apparatus of claim 8 , wherein the processor obtains the CT image pair for pediatrics, the CT image pair including the basic CT image and the suppression CT image for the pediatrics.
12. The apparatus of claim 8 , wherein the processor applies the style conversion model to the CT image pair for pediatrics to convert the CT image pair for the pediatrics into the conversion image pair for the pediatrics.
13. The apparatus of claim 8 , wherein the processor uses a basic conversion image of the conversion image pair of the training data as a training input image and uses a suppression conversion image of the conversion image pair as a ground truth output image to train the bone suppression model by means of supervised learning.