3D Medical Image Domain Conversion With 2D GAN Discriminators
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods struggle to generate high-resolution three-dimensional medical images across different modalities due to the challenge of training on low-resolution data in some axis directions, leading to difficulties in domain conversion for three-dimensional images.
Innovation Solution
A method involving a machine learning system with a three-dimensional generator and two-dimensional discriminators is employed, utilizing isotropic and cutout processing to train on low-resolution data, ensuring high-resolution generation across all axis directions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a learned model is created by a radiologist reading a large number of CT images to understand disease development, then the model accuracy is improved, but the time and cost required for model creation increases
Solution Approach 1:
The patent uses GANs to generate synthetic CT images that copy the characteristics of real medical images. Instead of requiring radiologists to manually analyze numerous real images, the system creates artificial training data that replicates disease progression patterns, thereby maintaining model accuracy while dramatically reducing the time and expertise required for data preparation
Solution Approach 2:
The system performs preliminary data preparation by generating synthetic training images before the actual model training begins. This advance preparation of training data eliminates the need for radiologists to manually curate and annotate large datasets, resolving the contradiction between achieving high model accuracy through extensive data review and the time cost of such manual processes
2Reliability
If real patient data is used for training, then the model reflects actual disease progression, but patient privacy and data security concerns arise
Solution Approach 1:
The patent employs GANs to create synthetic copies of real patient data that preserve the statistical properties and disease progression patterns of actual medical images without containing any real patient information. This copying approach maintains model reliability by ensuring the training data reflects true disease dynamics while eliminating privacy risks associated with using real patient data
Solution Approach 2:
The GAN-based synthetic data generation system acts as an intermediary between real patient data and the training model. Rather than directly using sensitive patient information, the system creates an intermediate synthetic dataset that mediates the training process, preserving disease realism while protecting patient privacy through mathematical transformation rather than direct data usage
3Adaptability or versatility
If more training data is collected to improve model generalization, then the model performance on unseen data improves, but the complexity of data management and processing increases
Solution Approach 1:
The patent uses GANs to generate synthetic training data that copies the diversity and distribution characteristics of real medical images. This approach improves model generalization by providing abundant varied training examples without the complexity of collecting, storing, and managing large volumes of real patient data, as the synthetic data can be generated on-demand with controlled diversity
Solution Approach 2:
The system performs preliminary generation of diverse synthetic training data with controlled variability to improve model generalization. This advance preparation of varied training examples eliminates the need for complex ongoing data collection and management processes, resolving the contradiction between achieving broad model adaptability and maintaining simple data management
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Provided are a method of generating a trained model, a machine learning system, a program, and a medical image processing apparatus capable of generating a high resolution three-dimensional generated image by converting a domain of a three-dimensional image which is input. A method of generating a trained model according to an aspect of the present disclosure, in which a learning model is used, which has a structure of a generative adversarial network including a first generator configured using a three-dimensional convolutional neural network that receives an input of a three-dimensional image of a first domain and that outputs a three-dimensional generated image of a second domain different from the first domain, and a first discriminator configured using a two-dimensional convolutional neural network that receives an input of a two-dimensional image indicating a cross section image in a first slice plane direction cut out from the three-dimensional generated image of the second domain and that discriminates authenticity of the input two-dimensional image, and the method includes, by a computer, performing training processing of training the first generator and the first discriminator in an adversarial manner based on a plurality of pieces of training data including a three-dimensional image captured under a first imaging condition and a three-dimensional image captured under a second imaging condition different from the first imaging condition.