Data-driven elasticity imaging using neural network constitutive models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional quasi-static elasticity imaging techniques require assumptions about tissue mechanical properties and internal structure, which can lead to inaccurate results due to measurement noise and limited data sampling, and often assume linear, isotropic, and incompressible tissues, failing to accurately model the nonlinear and viscoelastic properties of biological tissues.
Innovation Solution
The use of Cartesian neural network constitutive models (CaNNCMs) that learn stress and strain distributions from force-displacement data without prior knowledge of the tissue structure, allowing for the estimation of mechanical properties and geometric information independently of the internal structure, using a data-driven approach with artificial neural networks and the Autoprogressive Algorithm.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional model-based techniques are used to estimate mechanical parameters from force-displacement measurements, then the inverse problem can be solved with available data, but the tissue properties are incorrectly constrained to linear-elastic, isotropic, and incompressible assumptions that do not reflect actual biological tissue behavior
Solution Approach 1:
The patent transforms the constitutive model from fixed parametric assumptions to a data-driven neural network that learns stress-strain relationships directly from force-displacement measurements. The neural network adapts its parameters (weights and biases) during training to capture the actual nonlinear and viscoelastic behavior of biological tissues, replacing conventional linear-elastic models with a flexible functional form that can represent complex material properties without predefined constraints
Solution Approach 2:
The patent substitutes the conventional mechanical constitutive model (based on physical assumptions about tissue behavior) with an informational model (neural network) that learns the mechanical behavior from data. This replacement allows the system to capture complex nonlinear and viscoelastic properties without relying on simplified mechanical assumptions, effectively using information processing to replace traditional mechanical modeling
2Ease of manufacture
If simplifying assumptions are imposed to overcome the ill-posed nature of the inverse problem, then the problem becomes solvable with limited data, but parametric errors are made that corrupt the final elastogram
Solution Approach 1:
The patent performs preliminary training of the neural network constitutive model using force-displacement data before solving the inverse problem. This pre-training phase allows the model to learn the stress-strain relationship and material properties from available measurements, preparing an accurate constitutive model that can then be used to solve the inverse problem without requiring simplifying assumptions about tissue behavior
Solution Approach 2:
The patent introduces a neural network constitutive model as an intermediary between force-displacement measurements and the final elastogram. This intermediary learns the complex stress-strain relationship from data and provides accurate material property estimates that bridge the gap between limited measurements and the complete mechanical property distribution, avoiding the need for direct inversion with simplifying assumptions
3Reliability
If a chosen constitutive model is appropriate for certain tissue types, then accurate results are obtained for those tissues, but the model becomes incorrect for other tissues in the same field of view, leading to model selection errors
Solution Approach 1:
The patent creates a universal neural network constitutive model that can accurately represent multiple tissue types within the same field of view. The neural network learns the stress-strain relationship from force-displacement data without being constrained to a specific tissue type, making it applicable to heterogeneous tissues with different mechanical properties. This single model replaces the need for separate constitutive models for different tissue types
Solution Approach 2:
The patent transforms the static constitutive model into a dynamic, data-driven model that adapts to different tissue types through learning. The neural network parameters are adjusted during training based on the actual force-displacement measurements, allowing the model to dynamically adapt to the mechanical properties of whatever tissues are being imaged, rather than being fixed to represent a specific tissue type
Data Source
AI summary
Systems and methods are provided for employing informational models trained using the Autoprogressive Algorithm to learn the mechanical behavior and internal structure of biological materials using a sparse sampling of force and displacement measurements. Forces are applied to the biological material and the force and displacement are measured and applied to the AutoP algorithm. Additionally, a coordinate based scaling factor is applied to the measured displacement.