Patient-Specific Breast Model for DBT View Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for ipsilateral mapping of Digital Breast Tomosynthesis (DBT) views, such as those using hemispherical compression models, face challenges with computational complexity and insufficient mapping accuracy, making it difficult to quickly and accurately align suspicious areas between different breast scan orientations.
Innovation Solution
A method employing patient-specific breast modeling using machine learning-based data-driven approaches to intermediate map localisation data from one DBT view to another, eliminating the need for biomechanical modeling and GPU implementation, allowing for precise and efficient alignment of first and second localisation data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If hemispherical compression models are used for ipsilateral mapping, then mapping can be performed, but computational complexity increases and mapping accuracy becomes insufficient
Solution Approach 1:
The patent replaces complex biomechanical modeling and simulation systems with a simplified data-driven approach. Instead of using hemispherical compression models that require GPU implementation and complex calculations, the invention uses machine learning models trained on breast image data to directly predict corresponding locations between MLO and CC views, significantly reducing computational complexity while maintaining or improving mapping accuracy
Solution Approach 2:
The patent changes the approach from physical model-based parameter estimation to data-driven parameter prediction. By training machine learning models on actual breast image data and corresponding locations, the system learns optimal mapping parameters directly from data rather than deriving them from simplified hemispherical assumptions, improving accuracy without increasing computational burden
2Measurement precision
If complex biomechanical modeling is used for mapping, then mapping can be achieved, but the process requires GPU implementation and lengthy computation time
Solution Approach 1:
The patent performs preliminary action by pre-training machine learning models on large datasets of breast images and corresponding locations before actual mapping is needed. This offline training phase captures the complex mapping relationships, so that during actual clinical use, only fast inference is required. This eliminates the need for lengthy computation during patient workflow while maintaining high precision
Solution Approach 2:
The patent substitutes complex biomechanical simulation systems with trained machine learning models. The complex physics-based calculations that require GPU implementation are replaced with lightweight predictive models that can perform mapping in real-time during clinical workflows, dramatically reducing computation time while preserving mapping precision
3Productivity
If simplified hemispherical models are used for mapping, then computational effort is reduced, but mapping accuracy becomes insufficient
Solution Approach 1:
The patent changes from using fixed geometric parameters of hemispherical models to using data-driven parameters learned from actual breast images. The machine learning models learn optimal mapping parameters specific to each patient's breast anatomy and compression characteristics, achieving high accuracy without the computational overhead of complex biomechanical models
Solution Approach 2:
The patent creates accurate copies of the complex mapping relationships through machine learning models trained on paired MLO and CC breast images. Instead of using simplified hemispherical approximations, the system learns and replicates the actual deformation patterns from training data, achieving high fidelity mapping with efficient inference
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
The present invention concerns a method (Z) of assigning first localisation data (LD1) of a breast (4) of a patient (P) derived in first image data (ID1) of the breast (4), the first image data (ID1) being the result of a first radiological data acquisition process, to second localisation data (LD2) of the same breast (4) derived in second image data (ID2), the second image data (ID2) being the result of a second radiological data acquisition process, or vice versa. Thereby, the first localisation data (LD1) are assigned to the second localisation data (LD2) by intermediately mapping them into breast model data (MD) representing a patient-specific breast shape of the patient (P) and then onto the second image data (ID2) - or vice versa, thereby deriving assignment data (AD1, AD2). The invention also concerns an assignment system (33) for such purpose.