Synthetic CMR Data Generation via Parameter Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current synthetic data generation for cardiac magnetic resonance (CMR) imaging lacks control over acquisition parameters, resulting in images that may appear realistic but lack complexity, which can lead to suboptimal deep learning model performance due to unbalanced or small datasets.
Innovation Solution
A machine-learned model generates synthetic CMR imaging samples by inputting values for parameters such as electrocardiogram (ECG), image style, number of slices, pathology, and heart function, using a latent representation to enhance control and variance, with the model being a generative adversarial network or autoencoder trained on ground truth images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If synthetic data is generated using current approaches, then the dataset size increases, but the control over acquisition parameters is insufficient and the images lack complexity
Solution Approach 1:
The patent applies parameter changes by systematically varying acquisition parameters (magnet strength, field of view, slice thickness, echo time, repetition time, flip angle) as inputs to the generative model. This enables control over the complexity and characteristics of generated synthetic images while maintaining realistic appearances, directly resolving the contradiction between dataset quantity and parameter control versatility.
2Reliability
If more annotated examples are acquired to improve deep learning performance, then the model performance improves, but the acquisition cost and time increase
Solution Approach 1:
The patent employs copying by using a generative adversarial network to create synthetic copies of medical images that mimic real annotated examples. These synthetic copies serve as training data for deep learning models, providing the necessary quantity and quality of training examples without the time-consuming process of acquiring and annotating real patient data, thus improving model performance while reducing time loss.
3Adaptability or versatility
If the training dataset is enlarged to cover more situations, then the model generalization improves, but the data acquisition becomes even more difficult
Solution Approach 1:
The patent implements self-service by enabling the synthetic data generation system to automatically produce diverse training examples covering various medical situations and pathologies without manual intervention. The system self-generates images with different acquisition parameters, anatomical variations, and pathological conditions, making the data acquisition process effortless while improving model generalization across diverse scenarios.
Data Source
AI summary
CMR imaging is synthesized, and/or machine learning for a CMR imaging task uses synthetic sample generation. A machine-learned model generates synthetic samples. For example, the machine-learned model generates the synthetic samples in response to input of values for two or more parameters from the group of electrocardiogram (ECG), an indication of image style, a number of slices, a pathology, a measure of heart function, sample image, and/or an indication of slice position relative to anatomy. The indication of image style may be in the form of a latent representation, which may be used as the only input or one of multiple inputs. These inputs provide for better control over generation of synthetic samples, providing for greater variance and breadth of samples then used to machine train for a CMR task.