Method of generating training data and medical image processing apparatus for deep-learning-based mammography denoising
Systems and methods are provided for generating training data for training a neural network. The neural network may be a machine learning model. The neural network may be a U-Net neural network. Training data for training a neural network may be generated by combining X-ray image data and non-X-ray image data. Training data for training a neural network may be generated by superimposing data having a characteristic related to a lesion onto X-ray image data. The superimposed data may be used to simulate calcification features in mammographic images. Processing circuitry is configured to perform noise reduction by inputting an input image of a breast into a trained machine-learning model to output an output image in which noise included in the input image is reduced in the output image while leaving a characteristic of a lesion included in the input image undeleted in the output image.
1 . A method of generating training data for training a machine-learning model, the method comprising:
obtaining first X-ray image data acquired by irradiating with X-rays a breast of an object, the first X-ray image data including an X-ray image of the breast;
obtaining photographic image data, wherein the photographic image data does not include a photographic image of a breast;
generating combined image data based on the first X-ray image data and the photographic image data, the combined image data being generated by combining at least a portion of the first X-ray image data with at least a portion of the photographic image data;
generating calcification signal data for simulating calcification; and
generating training data for training a machine-learning model for outputting information based on second X-ray image data, the training data being generated by adding the calcification signal data to the combined image data.
2 . The method of claim 1 , wherein the combined image data including the calcification signal data is used as target data in training the machine-learning model.
3 . The method of claim 2 , wherein the trained machine-learning model is used for generating a noise-reduced image corresponding to the second X-ray image data.
4 . The method of claim 1 , wherein in generating the combined image data, a contour line of the breast and a background component in the first X-ray image data and a textural feature in the photographic image data are combined.
5 . The method of claim 1 , wherein the photographic image data includes a feature that simulates one or more features characteristic of a mammographic image.
6 . The method of claim 1 , wherein generating the training data includes superimposing data having a characteristic related to a lesion onto the combined image data, the data having the characteristic related to the lesion comprising the calcification signal data.
7 . A medical image processing apparatus comprising:
processing circuitry configured to perform noise reduction by inputting an input image of a breast into a trained machine-learning model to output an output image in which noise included in the input image is reduced in the output image while leaving a characteristic of a lesion included in the input image undeleted in the output image,
wherein the trained machine-learning model, for reducing noise in an image while leaving a characteristic of a lesion included in the image undeleted, is generated by training a machine-learning model with a first image as training data for an output side of the machine-learning model, and a second image as training data for an input side of the machine-learning model,
the first image is generated by:
generating combined image data based on first image data including an image of a breast of an object and second image data not including an image of a breast, the combined image data being generated by combining at least a portion of the first image data with at least a portion of the second image data,
generating calcification signal data for simulating calcification, and
adding the calcification signal data to the combined image data to generate the first image,
the first image includes a characteristic of the lesion, and
the second image includes higher noise than the first image and includes the characteristic of the lesion included in the first image.
8 . The medical image processing apparatus of claim 7 , wherein the training includes identifying a region outside a breast region in the training data and determining whether training is finished based on information corresponding to the breast region.
9 . The medical image processing apparatus of claim 8 , wherein the region outside the breast region is identified based on values of pixels in the region outside the breast region exceeding a threshold.
10 . The medical image processing apparatus of claim 8 , wherein the determining whether the training is finished is based on loss values calculated with respect to the breast region.
11 . The medical image processing apparatus of claim 7 , wherein the image of the breast of the object is an X-ray image.
12 . The medical image processing apparatus of claim 7 , wherein the first image data is X-ray image data, and the image of the breast of the object is an X-ray image, and
wherein the second image data is photographic image data, and the photographic image data does not include a photographic image of a breast.
13 . The medical image processing apparatus of claim 7 , wherein the second image is generated by adding noise to the first image,
wherein the first image data is X-ray image data, and
wherein the second image data is photographic image data.