Medical Image Characterization Using LDDMM and Sparse Anomaly Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning techniques for characterizing medical images require large labeled training sets and are prone to overfitting, failing to generalize well on new data.
Innovation Solution
A method using Large Diffeomorphic Deformation Metric Mapping (LDDMM) to create a template from a population of anomaly-free images, minimizing a cost function to determine deformation and anomaly information without labels, allowing for efficient characterization of new images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning techniques are used to characterize medical images, then anomaly detection capability is improved, but the requirement for large labeled training sets increases and overfitting occurs
Solution Approach 1:
The patent extracts and removes the requirement for large labeled training sets from the anomaly detection process. By formulating the problem as finding a sparse anomaly matrix A that explains deviations from a learned template, the method separates anomaly detection from traditional supervised learning requirements, achieving effective anomaly detection without needing extensive labeled training data
Solution Approach 2:
The patent changes the fundamental parameters of the approach by shifting from supervised learning with labeled data to an unsupervised framework based on template matching and sparsity assumptions. This parameter change in the methodological approach eliminates the need for large training sets while maintaining anomaly detection capability
2Measurement precision
If deep learning techniques are used to characterize medical images, then anomaly detection capability is improved, but overfitting occurs and generalization to new data fails
Solution Approach 1:
The patent enables the system to serve itself by automatically learning the template and anomaly characteristics from the data without requiring external labeled training sets. The sparsity constraint on the anomaly matrix provides built-in regularization that prevents overfitting, allowing the method to generalize well to new data while maintaining high anomaly detection capability
Solution Approach 2:
The patent introduces a learned template as an intermediary representation that mediates between the input images and anomaly detection. This template serves as a reference model that captures normal variations, allowing the system to generalize to new data by comparing against this intermediate representation rather than relying on memorized training examples
3Measurement precision
If traditional deep learning methods are used, then training accuracy is improved, but the complexity of labeling and annotation increases
Solution Approach 1:
The patent extracts and removes the annotation step from the process entirely. By formulating anomaly detection as a sparse recovery problem where the anomaly matrix A is learned directly from image deviations from a template, the method eliminates the need for manual labeling and annotation while maintaining training accuracy through the sparsity constraint
Data Source
AI summary
A method implemented by a computer for characterizing at least one observation y of a subject, including the steps of determining a template y characterizing a population of subjects without anomaly, by diffeomorphic deformation, minimizing the cost function J: J(F,A)=K∥y−F(y)−A∥2+λ∥A∥1+Reg(F) where F is a deformation function, A is an anomaly matrix, ∥A∥1 is the 1-norm of A, K and λ are predefined constants, Reg is a regularization function, where during minimization, F and A are determined by learning, F providing information on the morphological variability of the subject and A providing information on anomalies in the observation y.