Simulating Contrast Agent Dose Images for Medical Imaging Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical imaging techniques require prospective studies with dedicated protocols to collect sample sets for training deep learning models, which are costly, risky, and time-consuming, limiting the quality and diversity of training data and impacting the robustness of the models, leading to potential errors in diagnostic and therapeutic applications.
Innovation Solution
A method for training machine learning models by simulating source images from baseline and target images, allowing incomplete sample sets to be completed without requiring actual acquisition of all dose levels, using a neural network to optimize the generation of target images from baseline and source images, with the option to simulate higher contrast doses for enhanced imaging.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If prospective studies with dedicated protocols are conducted to collect sample sets for training deep learning models, then the quality and diversity of training data is improved, but the cost, time consumption, and risk increase significantly
Solution Approach 1:
The patent uses image simulation technology to generate virtual source images that copy the characteristics of actual medical images without requiring physical acquisition. This allows the training system to use simulated images as substitutes for real patient images, thereby avoiding the need for prospective studies while maintaining data quality and diversity for model training
Solution Approach 2:
The system performs self-service by generating its own training data through simulation. Instead of relying on external prospective studies to collect sample sets, the system autonomously creates synthetic source images from available baseline and target images, eliminating the time-consuming data collection process while maintaining sufficient training data quality
2Reliability
If prospective studies with dedicated protocols are conducted to collect sample sets for training deep learning models, then the quality and diversity of training data is improved, but the cost and risk increase significantly
Solution Approach 1:
The patent creates virtual copies of medical images through simulation rather than acquiring real patient images. These synthetic source images replicate the essential characteristics needed for training without exposing actual patients to additional risks, thereby maintaining training data quality while eliminating patient harm
Solution Approach 2:
The patent converts the limitation of not having access to diverse real patient data into a benefit by using simulation to generate synthetic data. This approach transforms the potential harm of requiring prospective studies into a beneficial method that avoids patient risk while still providing sufficient training data diversity
3Quantity of substance
If reduced-dose images are acquired to train the deep learning network, then the contrast agent dose is reduced, but additional images must be acquired which may expose patients to unneeded radiations
Solution Approach 1:
The patent generates virtual reduced-dose images through simulation rather than acquiring actual reduced-dose images from patients. This copying approach allows the system to train on diverse dose levels without exposing patients to additional radiation, as the simulated images replicate the appearance of reduced-dose acquisitions without the associated radiation risk
Solution Approach 2:
The patent introduces image simulation technology as an intermediary between the need for reduced-dose training data and patient safety. This intermediary layer allows the system to obtain virtual reduced-dose images for training purposes without directly exposing patients to radiation, thereby mediating between data quality requirements and patient protection
Data Source
AI summary
A solution is proposed for training a machine learning model (420) for use in medical imaging applications. A corresponding method (700) comprises providing (703-743: 759-763) sample sets, each comprising a sample baseline image, a sample target image (acquired from a corresponding body-part of a subject to which a contrast agent at a certain dose has been administered) and a sample source dose (corresponding to a different dose of the contrast agent). The machine learning model (420) is trained (744-758) so as to optimize its capability of generating each sample target image from the corresponding sample baseline image and sample source image. One or more of the sample sets are incomplete, missing their sample source images. Each incomplete sample set is completed (704-742: 759-763) by simulating the sample source image from at least the sample baseline image and the sample target image of the sample set. A computer programs (500) and a computer program products for implementing the method (700) are proposed. Moreover, a computing system (130) for performing the method (700) is proposed.