Statistical Anatomical Model Segmentation via Physics-Based Simulated Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical imaging techniques face challenges in anatomical segmentation, particularly in lower quality imaging modalities like x-ray or ultrasound, due to high intra-observer and inter-observer variability and limited image content, which leads to errors in model training and ambiguity.
Innovation Solution
A computer-implemented method that creates a statistical anatomical model from high-resolution imaging data, generates simulated images based on physics principles, and uses a matching algorithm to deform the model until it matches an unlabeled input image, allowing for accurate segmentation across various imaging modalities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing statistical models (active shape models, active appearance models) are used for anatomical segmentation, then segmentation can be performed, but the models are restricted to the modality in which they are trained and generate errors in low-quality modalities due to intra-observer and inter-observer variability
Solution Approach 1:
The patent creates a 3D anatomical model from high-resolution training data (CT/MRI) and generates simulated 2D images from this model that replicate the appearance of low-quality modalities (X-ray, ultrasound). This copying approach allows the model to adapt to different imaging modalities without retraining, maintaining segmentation accuracy across modalities by synthesizing representative images from the high-resolution source model.
Solution Approach 2:
The patent introduces an intermediary simulation step that translates the 3D anatomical model into simulated 2D images matching the target modality's appearance characteristics. This intermediary simulation acts as a bridge between the high-resolution training data and the low-quality input images, enabling the statistical model to handle modalities it wasn't directly trained on while maintaining segmentation reliability.
2Adaptability or versatility
If models are trained on low-quality imaging modalities, then the model can process that modality, but image ambiguity and limited content lead to increased error and variability
Solution Approach 1:
The patent performs preliminary action by training the statistical model on high-resolution imaging data (CT/MRI) where anatomical structures are clearly visible and unambiguous. This preliminary training on high-quality data establishes accurate anatomical representations before the model encounters low-quality modalities during inference, allowing it to maintain high segmentation precision even when processing ambiguous low-quality images.
Solution Approach 2:
The patent generates simulated images from the 3D anatomical model that copy the appearance characteristics of the target low-quality modality. These simulated images serve as training examples that preserve the anatomical accuracy of high-resolution modalities while replicating the appearance artifacts and characteristics of low-quality modalities, enabling the model to achieve both modality coverage and segmentation precision.
3Measurement precision
If detailed high-resolution imaging data is used for training, then anatomical detail and accuracy are improved, but the data is not directly applicable to low-quality modalities due to modality-specific characteristics
Solution Approach 1:
The patent creates a 3D anatomical model from high-resolution training data and then generates simulated 2D images from this model that copy the appearance characteristics of various imaging modalities. This allows the model to retain the anatomical detail accuracy of high-resolution modalities while adapting to the appearance characteristics of low-quality modalities through the simulation process, achieving both precision and cross-modality applicability.
Solution Approach 2:
The patent changes the parameter representation by moving from direct 2D image analysis to a 3D anatomical model representation. This parameter change allows the model to capture anatomical structures in three dimensions with high precision, then project these structures into 2D simulated images that match the appearance characteristics of different modalities, thereby achieving both anatomical accuracy and cross-modality adaptability.
Data Source
AI summary
Systems and articles of manufacture for image segmentation are provided herein, and include creating an anatomical model from training data comprising one or more imaging modalities, generating one or more simulated images in a target modality based on the anatomical model and one or more principles of physics pertaining to image contrast generation, and comparing the one or more simulated images to an unlabeled input image of a given imaging modality to determine a simulated image of the one or more simulated images to represent the unlabeled input image.