Apparatus and method for medical image denoising based on deep learning

1. A method for medical image denoising based on deep learning, the method comprising: generating multiple trained deep learning models, the multiple trained deep learning models being grouped by examination areas; extracting examination information from an input CT data, the examination information including examination area information; selecting at least one deep learning model corresponding to the examination information from the multiple trained deep learning models; and outputting a CT data denoised from the input CT data by feeding the input CT data into the selected at least one deep learning model wherein the generating comprises: generating a second training CT data set to which noises of multiple predetermined levels are added by applying a CT data image noise simulator to a first training CT data set; extracting examination information from the second training CT data set and grouping the second training CT data set into multiple groups according to a predetermined rule; and generating and training multiple training-target deep learning models so as to correspond to the respective groups of the second training CT data set by groups, wherein in the selecting, the multiple previously trained deep learning models are the multiple training-target deep learning models trained in the generating and training.

2. The method of claim 1 , wherein in the generating of the second training CT data set, the CT data noise simulator performs, generating a synthetic sinogram by inputting an image of the first training CT data set; generating a noise-component synthetic sinogram by applying a noise model having a predetermined level of noise to the synthetic sinogram; generating a noise-component CT data based on the generated noise-component synthetic sinogram; and generating a synthetic high noise CT data by adding the generated noise-component CT data to the original CT data.

3. The method of claim 1 , wherein the second training CT data set is composed of a pair of synthetic high noise CT data set and synthetic noise-component CT data set obtained on the basis of the first training CT data set.

4. The method of claim 1 , wherein the generating and training includes feeding the synthetic high noise CT data set for the respective groups of the second training CT data set into the input of the corresponding training-target deep learning model and training the training-target deep learning model repeatedly by groups to minimize the difference between the synthetic noise-component CT data and output from the training-target deep learning model, in order for the multiple training-target deep learning model to have a function of extracting a noise-component CT data image from an input CT data.

5. The method of claim 1 , wherein the outputting includes feeding the input CT data into the selected at least one deep learning model and forcing the selected at least one deep learning model to extract a noise-component CT data from the input CT data and producing the denoised CT data by multiplying the extracted noise-component CT data with a predetermined value and subtracting the result from the input CT data.

6. An apparatus for medical image denoising based on deep learning, the apparatus comprising: an extraction unit configured to extract examination information from an input CT data, the examination information including examination area information; a selection unit configured to select at least one deep learning model corresponding to the examination information from multiple previously trained deep learning models; an output unit configured to output a CT data denoised from the input CT data by feeding the input CT data into the selected at least one deep learning model; and a training unit configured to generate and train the multiple previously trained deep learning models, the multiple previously trained deep learning models being grouped by examination areas wherein the training unit generates a second training CT data set to which noises of multiple predetermined levels are added by applying a CT data noise simulator to a first training CT data set, extracts examination information from the second training CT data set and groups the second training CT data set into multiple groups according to a predetermined rule, and generates and trains multiple training-target deep learning models so as to correspond to the respective groups of the second training CT data set by groups, and the multiple previously trained deep learning models are the multiple training-target deep learning models trained by the training unit.

7. The apparatus of claim 6 , wherein the CT data noise simulator generates a synthetic sinogram by inputting an image of the first training CT data set, generates a noise-component synthetic sinogram by applying a noise model having a predetermined level of noise to the synthetic sinogram, generates a noise-component CT data based on the generated noise-component synthetic sinogram, and generates a synthetic high noise CT data by adding the generated noise-component CT data to the original CT data.

8. The apparatus of claim 6 , wherein the second training CT data set is composed of a pair of synthetic high noise CT data set and synthetic noise-component CT data set obtained on the basis of the first training CT data set.

9. The apparatus of claim 6 , wherein the training unit feeds synthetic high noise CT data set for the respective groups of the second training CT data set into the input of the corresponding training-target deep learning model and trains the training-target deep learning model repeatedly by groups to minimize the difference between the synthetic noise-component CT data and output from the training-target deep learning model, in order for the multiple training-target deep learning model to have a function of extracting a noise-component CT data from an input CT data.

10. The apparatus of claim 6 , wherein the output unit feeds the input CT data into the selected at least one deep learning model and forces the selected at least one deep learning model to extract a noise-component CT data from the input CT data and produces the denoised CT data by multiplying the extracted noise-component CT data with a predetermined value and subtracting the result from the input CT data.

11. A non-transitory computer-readable storage medium that stores a program configured to implement a method according to claim 1 on a computer.