Trainable Function for Synthetic Medical Image Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Comparing medical image data from different modalities and protocols is limited due to poor registration, leading to unsuitable or limited synthetic image data, which increases costs and time without providing an unnecessary dose to the examination object.
Innovation Solution
A method using a trainable function, such as a Generative Adversarial Network, to determine synthetic image data by registering and transforming medical image data from one modality or protocol to another, optimizing parameters for improved similarity and registration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If medical image data is acquired with different medical modalities and imaging protocols to improve comparability, then the presentation and diagnostic value are improved, but the examination object is subjected to unnecessary additional radiation dose and acquisition costs increase
Solution Approach 1:
The patent creates synthetic image data that copies the presentation characteristics of a second medical imaging modality from first medical image data. Instead of acquiring actual second modality images (which would expose the patient to additional radiation), the system generates synthetic copies that mimic the appearance and diagnostic features of the target modality, thereby resolving the contradiction between obtaining comparable multi-modality images and avoiding unnecessary radiation exposure
Solution Approach 2:
The patent replaces the physical mechanical process of acquiring additional medical images with a computational image synthesis process. Rather than physically exposing the patient to another imaging modality (mechanical/physical acquisition), the system uses trained neural networks to computationally transform existing images into synthetic images of the desired modality, eliminating the need for additional physical exposure while maintaining diagnostic comparability
2Reliability
If medical image data is re-acquired with alternate modalities to improve comparability, then diagnostic quality is improved, but time consumption and acquisition costs increase
Solution Approach 1:
The patent performs preliminary actions by acquiring and processing medical image data from one modality that can subsequently be transformed into multiple other modalities through synthetic image generation. Instead of performing multiple sequential acquisitions (which would consume time), the system prepares the first modality images in advance and then rapidly generates synthetic versions of other modalities on-demand, significantly reducing the time required for multi-modality comparison while maintaining diagnostic quality
Solution Approach 2:
The system creates synthetic copies of medical images in different modalities from a single acquisition, eliminating the need for time-consuming re-acquisitions. The trained neural networks generate these synthetic copies rapidly, providing multiple modality presentations without the time penalty of performing separate physical scans, thus resolving the contradiction between diagnostic quality and time consumption
3Manufacturing precision
If registration is performed before determining synthetic image data to improve registration quality, then synthetic image data quality is improved, but the registration is limited by poor initial alignment of unregistered image data
Solution Approach 1:
The patent performs preliminary registration of the first medical image data with the second medical image data before using the registered data as input for synthetic image generation. This preliminary alignment ensures that when the neural network generates synthetic images, the spatial correspondence between source and target modalities is already optimized, allowing the synthesis process to focus on intensity and texture transformation rather than also solving the registration problem, thereby improving both registration accuracy and synthetic image quality
Solution Approach 2:
The patent merges the registration process with the synthetic image generation process by using registered image pairs as training data for the neural network. The system combines spatial alignment information from registration with the image transformation task, creating a unified approach where registration and synthesis work together synergistically to produce high-quality synthetic images with accurate spatial correspondence
Data Source
AI summary
A computer-implemented method includes receiving first medical image data, wherein the first medical image data is based on a first medical imaging of an examination object, receiving second medical image data, wherein the second medical image data is based on a second medical imaging of the examination object, wherein the first and the second medical imaging differ by at least one of an imaging modality or by an imaging protocol used, wherein the first and the second medical image data are registered with one another, determining synthetic image data by applying a trainable function to the first medical image data, determining a measure of similarity with a similarity function by comparison of the synthetic image data and the second medical image data, adjusting at least one parameter of the trainable function by optimization of the similarity function based on the measure of similarity, provision of the trainable function.