DE-112024002962-T5 - DEVICE AND METHOD FOR NEUTRALIZING STYLES OF MEDICAL IMAGES BASED ON DEEP LEARNING

DE112024002962T5DE 112024002962 T5DE112024002962 T5DE 112024002962T5DE-112024002962-T5

Abstract

According to the present invention, a method for deep learning-based neutralization of medical image styles, which generates neutralized images to be input into an artificial intelligence diagnostic support program, comprises: a medical image acquisition step to acquire medical images for external processing; and a neutralized image generation step to generate neutralized images by inputting the medical images to be processed into deep learning style neutralization models that have been previously trained to neutralize imaging properties, wherein a plurality of deep learning style neutralization models are provided such that each of the medical images to be processed is neutralized by a deep learning style neutralization model that corresponds to the imaging properties of the medical image to be processed.

Inventors

LEE AH YEONG
KIM JONG HYO

Assignees

CLARIPI INC

Dates

Publication Date: 20260507
Application Date: 20240712
Priority Date: 20230713

Claims (16)

A method for deep-learning-based style neutralization of medical images to generate a neutralized image for input into an artificial intelligence-based diagnostic support program, wherein the method comprises: receiving a medical image for processing from an external source; and generating a neutralized image by inputting the medical image for processing into a style-neutralization deep-learning model that has been previously trained to neutralize the imaging properties of the medical image for processing, where the style-neutralization deep-learning model comprises a variety of style-neutralization deep-learning models, and the style-neutralization deep-learning model that matches the imaging properties of the medical image to be processed performs the neutralization of the medical image to be processed.
Procedure according to Claim 1 , where each deep learning model comprises: an inverse transformation deep learning model trained on the basis of raw data acquired from a patient and a reconstruction image reconstructed by reflecting the imaging properties in the raw data, and configured to output the raw medical image data for processing when it receives the medical image for processing; and an imitation deep learning model trained on the basis of the raw data acquired from the patient and the reconstruction image reconstructed by reflecting the imaging properties in the raw data, and configured to output the neutralized image for the medical image when it receives the raw medical image data for processing.
Procedure according to Claim 2 , in which a parameter relating to the imaging properties of the medical image is changed in the deep learning model for inverse transformation and the deep learning model for imitation, and the parameter includes at least one of the following: tube voltage (kVp), tube current (mAs), detection quantum efficiency (DQE), noise, focal spots, compression force and post-processing method.
Procedure according to Claim 2 , further comprising: when training the deep learning model for inverse transformation, obtaining raw data from a patient; obtaining raw data in which the imaging properties are reflected by reflecting the imaging properties in the raw data; obtaining a reconstruction image in which the imaging properties are reflected by reconstructing the raw data to reflect the imaging properties; and training the deep learning model for inverse transformation by pairing the raw data and the reconstruction image to output the raw data of the medical image for processing when the medical image is input into the deep learning model for inverse transformation for processing.
Procedure according to Claim 2 , further comprising: when training the deep learning model to imitate; obtaining raw data from a patient; obtaining raw data in which imaging properties are reflected by reflecting the imaging properties in the raw data; obtaining a reconstruction image in which the imaging properties are reflected by reconstructing the raw data to reflect the imaging properties; and training the deep learning model to imitate by pairing the raw data and the reconstruction image to output a neutralized image of the medical image for processing when the raw medical image data is input into the deep learning model for processing.
Procedure according to Claim 2 , in which both the deep learning model for inverse transformation and the deep learning model for imitation have a variety of deep learning models that differ in terms of the imaging properties to be converted.
Procedure according to Claim 6 , in which the deep learning model for inverse transformation comprises: a first deep learning model for inverse transformation trained on the basis of the raw data obtained from the patient and a reconstruction image reconstructed by reflecting a first imaging property in the raw data; and a second deep learning model for inverse transformation trained on the basis of the raw data obtained from the patient and a reconstruction image reconstructed by reflecting a second imaging property in the raw data was reconstructed, and the deep learning model for imitation features: a first deep learning model for imitation trained on the basis of the raw data obtained from the patient and a reconstruction image reconstructed by reflecting a first imaging characteristic in the raw data; and a second deep learning model for imitation trained on the basis of the raw data obtained from the patient and a reconstruction image reconstructed by reflecting a second imaging feature in the raw data.
Procedure according to Claim 7 , in which the style-neutralizing deep learning model features: a first style-neutralizing deep learning model, which is connected to the first inverse transformation deep learning model and the first imitation deep learning model; and a second style-neutralizing deep learning model, which is connected to the second inverse transformation deep learning model and the second imitation deep learning model.
Device for deep learning-based style neutralization of medical images, wherein the device includes a processing unit configured to produce a neutralized image that is input into an artificial intelligence-based diagnostic support program, where the processing unit is configured to: receive a medical image for processing from an external source; and produce a neutralized image by inputting the medical image for processing into a deep learning style neutralization model that has been previously trained to neutralize the imaging properties of the medical image for processing, where the deep learning style neutralization model includes a plurality of deep learning style neutralization models, and the deep learning style neutralization model that matches the imaging properties of the medical image for processing performs the neutralization of the medical image for processing.
Device according to Claim 9 , where each deep learning model comprises: an inverse transformation deep learning model trained on the basis of raw data acquired from a patient and a reconstruction image reconstructed by reflecting the imaging properties in the raw data, and configured to output the raw medical image data for processing when it receives the medical image for processing; and an imitation deep learning model trained on the basis of raw data acquired from the patient and the reconstruction image reconstructed by reflecting the imaging properties in the raw data, and configured to output the neutralized image for the medical image when it receives the raw medical image data for processing.
Device according to Claim 10 , in which a parameter relating to the imaging properties of the medical image is changed in the deep learning model for inverse transformation and the deep learning model for imitation, and the parameter has at least one of the following values: tube voltage (kVp), tube current (mAs), detection quantum efficiency (DQE), noise, focal spots, compression force and post-processing methods.
Device according to Claim 10 , in which raw data from a patient is obtained when training the deep learning model for inverse transformation; raw data reflecting the imaging properties is obtained by reflecting the imaging properties in the raw data; a reconstruction image reflecting the imaging properties is obtained by reconstructing the raw data taking the imaging properties into account; and the deep learning model for inverse transformation is trained by pairing the raw data and the reconstruction image to output the raw medical image data for processing when the medical image is input into the deep learning model for inverse transformation for processing.
Device according to Claim 10 , in which raw data from a patient are obtained when training the deep learning model to imitate the imaging properties; raw data reflecting the imaging properties are obtained by reflecting the imaging properties in the raw data; a reconstruction image reflecting the imaging properties is obtained by reconstructing the raw data taking the imaging properties into account; and the deep learning model is trained to imitate the imaging properties by pairing the raw data and the reconstruction image to generate a neutralized image of the to output medical image for processing when the raw data of the medical image is fed into the deep learning model for imitation.
Device according to Claim 10 , in which both the deep learning model for inverse transformation and the deep learning model for imitation have a multitude of deep learning models that differ from each other with respect to the mapping properties to be converted.
Device according to Claim 14 , where the deep learning model for inverse transformation comprises: a first deep learning model for inverse transformation trained on the basis of the raw data obtained from the patient and a reconstruction image reconstructed by reflecting a first imaging feature in the raw data; and a second deep learning model for inverse transformation trained on the basis of the raw data obtained from the patient and a reconstruction image reconstructed by reflecting a second imaging feature in the raw data, and the deep learning model for imitation comprises: a first deep learning model for imitation trained on the basis of the raw data obtained from the patient and a reconstruction image reconstructed by reflecting a first imaging feature in the raw data; and a second deep learning model for imitation, trained on the basis of the raw data obtained from the patient and a reconstruction image reconstructed by reflecting a second imaging property in the raw data.
Device according to Claim 15 , in which the style-neutralizing deep learning model features: a first style-neutralizing deep learning model, which is connected to the first inverse transformation deep learning model and the first imitation deep learning model; and a second style-neutralizing deep learning model, which is connected to the second inverse transformation deep learning model and the second imitation deep learning model.

Description

[Technical field] The disclosure relates to a device and a method for deep-learning-based neutralization of medical image styles, in particular a device and a method for deep-learning-based neutralization of medical image styles in order to neutralize image style properties (or image characteristics). [State of the art] Generally, medical devices such as X-ray machines, computed tomography (CT) scanners, and magnetic resonance imaging (MRI) scanners are used to acquire medical images. In modern medicine, the medical images acquired with such devices serve as a crucial basis for determining the presence and characteristics of lesions, thus informing decisions regarding the diagnosis and treatment of a patient. With recent advances in artificial intelligence technology, various technologies have been explored to support AI-based decision-making. For example, an AI-based diagnostic support technology has been developed in the Korean patent application no. 10-2108401 (with the title “IDENTIFICATION SERVER BY AN ARTIFICIAL INTELLIGENCE SYSTEM AND IMAGE PROCESSING SYSTEM BASED ON PACS THAT CONTAINS THIS”, granted on April 29, 2020). In this prior art, medical images acquired by various medical devices are fed into an artificial intelligence model to identify the presence of a lesion and, based on the identification result, to quickly initiate medical treatment. However, the imaging properties of images input into an AI-based diagnostic support program vary depending on the device type, imaging technique, etc. Therefore, if images with different imaging properties are input into the AI-based diagnostic support program, the problem arises that the diagnostic support performance varies depending on the images input. [Disclosure] [Technical problem] One aspect of the disclosure is to provide a device and method for deep learning-based neutralization of medical image styles, whereby input images with different imaging properties are standardized by neutralizing the imaging properties of the input images. [Technical solution] According to one embodiment of the disclosure, a method for deep-learning-based style neutralization of medical images to generate a neutralized image for input into an artificial intelligence-based diagnostic support program comprises the following: obtaining a medical image for processing from an external source; and generating a neutralized image by inputting the medical image for processing into a style-neutralization deep-learning model that has been previously trained to neutralize the imaging properties of the medical image for processing, wherein the style-neutralization deep-learning model comprises a plurality of style-neutralization deep-learning models, and the style-neutralization deep-learning model that matches the imaging properties of the medical image for processing performs the neutralization of the medical image for processing. Each deep learning model can include: an inverse transformation deep learning model trained on raw data acquired from a patient and a reconstruction image reconstructed by reflecting imaging properties in the raw data, configured to output the raw medical image for processing when it receives the medical image for processing; and an imitation deep learning model trained on the raw data acquired from the patient and the image reconstructed by reflecting imaging properties in the raw data, configured to output the neutralized image for the medical image when it receives the raw medical image for processing. In the deep learning model for inverse transformation and the deep learning model for imitation, a parameter relating to the imaging properties of the medical image can be used. refers to, can be changed, and the parameter can include at least one of the following parameters: tube voltage (kVp), tube current (mAs), detection quantum efficiency (DQE), noise, focal spots, compression force, and post-processing methods. When training the deep learning model for inverse transformation, the procedure may further include: obtaining raw data from a patient; obtaining raw data in which imaging properties are reflected by mirroring the imaging properties in the raw data; obtaining a reconstruction image in which the imaging properties are reflected by reconstructing the raw data to reflect the imaging properties; and training the deep learning model for inverse transformation by pairing the raw data and the reconstruction image to output the raw medical image data for processing when the medical image is input into the deep learning model for inverse transformation processing. When training the deep learning model for imitation, the procedure may further include: capturing raw data from a patient; capturing raw data in which the imaging properties are reflected by reflecting the imaging properties in the raw image data; obtaining a reconstruction image in which the imaging properties are reflected by reconstructing the raw data to reflect the imaging properties;