Method for generating trained model, machine learning system, program, and medical image processing device

The method addresses the challenge of converting low-resolution 3D medical images by using a generative adversarial network with 3D and 2D neural networks to enhance resolution, enabling high-quality image generation across varying imaging modalities.

JP7765415B2Active Publication Date: 2025-11-06FUJIFILM CORP
4 Cites 0 Cited by

Patent Information

Application Number
JP2022578244
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2021-01-26
Filing Date
2022-01-17
Publication Date
2025-11-06
Estimated Expiration
2042-01-17

AI Technical Summary

Technical Problem

Existing methods for domain conversion of medical images, particularly three-dimensional images, face challenges when training models due to the rarity of high-resolution data across different imaging modalities, leading to difficulties in generating high-resolution images when using low-resolution data, which is more readily available.

Method used

A method utilizing a generative adversarial network with a 3D convolutional neural network generator and 2D classifiers to convert 3D images from one domain to another, even with varying resolution conditions, by training on a mix of low-resolution data and using isotropic and anisotropic data processing to enhance resolution in specific directions.

Benefits of technology

Enables the generation of high-resolution 3D images across all cross-sectional directions, overcoming the limitations of using low-resolution data and varying imaging conditions, thereby expanding the applicability of AI-based diagnosis systems across different medical imaging modalities.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

Provided are a learned model generation method, a machine learning system, a program, and a medical image processing device with which a high-resolution 3D generated image can be generated by converting domains of an inputted 3D image. A learned model generation method according to one embodiment of the present disclosure includes performing learning processing in which a computer causes a first generator and a first discriminator to learn in an adversarial manner on the basis of a plurality of learning data items including a 3D image that is photographed under a first photography condition and a 3D image that is photographed under a second photography condition which is different from the first photography condition, such learning processing using a learning model that has the structure of a generative adversarial network including: the first generator, which is configured using a 3D convolutional neural network that receives an input of a 3D image of a first domain and that outputs a 3D generated image of a second domain that is different from the first domain; and the first discriminator, which is configured using a 2D convolutional neural network that receives an input of a 2D image representing a cross-sectional image in the direction of a first slice plane cut out from the second-domain 3D generated image, and that discriminates the authenticity of the inputted 2D image.
Need to check novelty before this filing date? Find Prior Art