Constrained Medical Image Data for Accurate Maskless Subtraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning methods, such as deep CNNs, used for predicting mask and subtraction volumes from contrast-enhanced data in medical imaging, often generate non-existent image features and struggle to distinguish between metal objects and enhanced vessels, leading to inaccuracies and the need for extensive testing to correct errors.
Innovation Solution
A medical image processing apparatus and method that utilize trained models to predict mask and subtraction data from contrast-enhanced medical image data, with the application of constraints through a constrained optimization procedure to refine the predictions and ensure accuracy, thereby reducing non-existent features and improving distinction between metal objects and enhanced vessels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If deep CNN is used to predict mask and subtraction volumes from contrast-enhanced data, then the prediction process is automated and speeded up, but non-existent image features are generated and metal objects cannot be distinguished from enhanced vessels
Solution Approach 1:
The patent implements a feedback mechanism where the predicted mask volume is used to generate a corrected subtraction volume by subtracting it from the contrast-enhanced volume. This feedback loop allows the system to identify and correct hallucinated features by ensuring that only features present in the original contrast volume appear in the final subtraction image, thereby improving prediction reliability while maintaining automation
Solution Approach 2:
The patent introduces an intermediary correction step between the initial CNN prediction and the final output. The predicted mask volume serves as an intermediary that mediates between the raw CNN prediction and the final subtraction volume, allowing problematic features to be identified and corrected before final output, thus resolving the contradiction between speed and accuracy
2Device complexity
If maskless DSA is implemented using trained models, then the need for separate mask volume acquisition is eliminated, but the models generate features that do not exist in reality requiring prohibitive model testing
Solution Approach 1:
The patent converts the harmful effect of hallucinated features into a beneficial correction process. By using the predicted mask volume to generate a corrected subtraction volume through subtraction from the original contrast volume, the system automatically identifies and removes hallucinated features without requiring extensive external testing, thus eliminating the need for separate mask acquisition while avoiding prohibitive testing time
Solution Approach 2:
The system performs self-correction by using its own predicted mask volume to identify and remove hallucinated features from the subtraction volume. This self-service mechanism eliminates the need for external validation and extensive model testing, maintaining the simplicity of maskless DSA while ensuring prediction accuracy
3Device complexity
If CNN predicts subtraction data directly from contrast volume, then the process is simplified to a single prediction step, but features present in contrast volume may disappear in predicted DSA and mask
Solution Approach 1:
The patent performs a preliminary action by first predicting the mask volume from the contrast-enhanced volume before generating the final subtraction volume. This preliminary mask prediction serves as a guide to identify which features should be preserved, ensuring that important anatomical structures are not lost in the subsequent subtraction process, thus maintaining feature preservation accuracy while keeping the overall process simplified
Data Source
AI summary
An apparatus for producing constrained medical image data, the apparatus including processing circuitry configured to: receive medical image data that includes or is obtained from scan data representing an anatomical region in which a sub-region is enhanced; predict, using a trained model, mask data from the medical imaging data, wherein the mask data is representative of the anatomical region without enhancement of the sub-region; and predict, using the trained model or a further trained model, subtraction data from the same medical image data, the subtraction data being representative of the same anatomical region, and the processing circuitry being further configured to apply at least one constraint to obtain constrained subtraction data.