ML Model Training for Metal Artefact Removal in Medical Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for removing metal and metal artefacts from medical images, such as CT and MRI scans, are inadequate due to limitations in algorithm applicability across different body parts, introduction of secondary artefacts, and the scarcity of training data for machine learning models.
Innovation Solution
A method and system for training a machine learning model using simulated images with metal artefacts, where the model is optimized to minimize differences between predicted and ground truth images, allowing for effective removal of metal and metal artefacts from medical scans.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If interpolation-based methods are used to replace metal projections, then metal artefacts can be reduced, but secondary artefacts are introduced
Solution Approach 1:
The patent uses simulated images as copies of real medical images, where synthetic metal artefacts are added to the simulated images to create training data. This allows the machine learning model to learn from numerous synthetic examples without requiring actual patient scans with metal implants, thus avoiding the introduction of secondary artefacts while still effectively reducing metal artefacts.
2Loss of time
If algorithms developed on one body part are applied to another, then development time is reduced, but algorithm performance deteriorates due to different body part characteristics
Solution Approach 1:
The patent creates a universal machine learning model through simulated training data that can be applied across different body parts. By training on diverse simulated images representing various anatomical regions, the model learns generalizable patterns that work across multiple body parts without requiring separate algorithm development for each region, thus maintaining both efficiency and performance.
3Object-affected harmful factors
If deep learning methods are used to learn artefact removal strategy, then removal capability is improved, but metal-corrupted regions are still not well recovered due to lack of training data
Solution Approach 1:
The patent performs preliminary action by pre-training the deep learning model using simulated images with synthetic metal artefacts before applying it to real medical images. This preliminary training on extensively generated synthetic data prepares the model to effectively handle metal-corrupted regions in actual clinical images, overcoming the limitation of insufficient real training data.
4Measurement precision
If medical CT images of the same subjects with and without metal artefacts are obtained for training, then training data quality is improved, but data acquisition becomes extremely difficult and limited
Solution Approach 1:
The patent creates copies of real medical images through simulation, adding synthetic metal artefacts to generate training pairs. This copying approach allows generation of unlimited training data without requiring actual re-scanning of patients with and without metal implants, thus maintaining high training data quality while eliminating the extreme difficulty of data acquisition.
Data Source
AI summary
A method and system for training a machine learning model for reducing or removing a foreign material or artefacts due to a foreign material from an image of a subject, the method comprising: generating one or more first simulated images from one or more real or simulated images of the foreign material (and optionally artefacts due to the foreign material), and from one or more real images of one or more subjects that are free of the foreign material and of artefacts due to the foreign material, such that the generated simulated images include the foreign material and artefacts due to the foreign material; generating one or more predicted images employing at least the first simulated images with a machine learning network that implements a machine learning model; and training or updating the machine learning model with the machine learning network by reducing or minimizing a difference between the one or more predicted images and ground truth data comprising one or more real or simulated images.