Data Augmentation via Point Distribution Model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data augmentation methods for training AI systems using computational models of body parts often result in unrealistic datasets due to random, non-systematic deformations, leading to poor performance in identifying features of interest.
Innovation Solution
A method is introduced to generate more realistic datasets by creating a point distribution model (PDM) based on an input dataset, using principal component analysis to define a surface model dataset with weight-eigenvector pairs, and modifying these weights to produce deformed volume datasets that are realistic and varied, ensuring accurate representation of body parts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If random deformations are applied to existing datasets to augment training data, then the number of training examples is increased, but the realism of the generated datasets deteriorates
Solution Approach 1:
The patent applies parameter changes by using statistical shape models that modify geometric parameters (weights of eigenvectors) in a systematic way. Instead of random deformations, the method changes shape parameters according to a statistical model derived from real patient data, generating variations that preserve anatomical realism while increasing dataset diversity.
Solution Approach 2:
The patent uses copying by creating synthetic patient models that copy the statistical characteristics of real patients. The statistical shape model captures the essential variations found in real patient anatomy, and synthetic models are generated by combining the mean shape with weighted eigenvector variations, effectively copying the statistical structure of real data without using actual patient information.
2Adaptability or versatility
If existing patient models are deformed in a random manner to create variations, then dataset diversity is increased, but the accuracy of representing real inter-patient differences deteriorates
Solution Approach 1:
The patent changes geometric parameters systematically using a statistical shape model. The model represents shape variations through a linear combination of eigenvectors with weights sampled from a probability distribution. This parameter change approach ensures that generated variations accurately reflect real inter-patient differences captured during model training, rather than introducing arbitrary random deformations.
Solution Approach 2:
The statistical shape model is trained on real patient data, creating a feedback loop where the model continuously refines its understanding of real anatomical variations. The eigenvectors and their associated weights are derived from analyzing actual patient measurements, ensuring that the generated variations are grounded in real medical data and accurately represent inter-patient differences.
Data Source
AI summary
A method for generating data representing the volume of part of a body, the method comprising generating a point distribution model “PDM” based on an input dataset comprising data representing at least one surface of part of a body, the PDM defining a surface model dataset based on an average dataset and one or more weight-eigenvector pairs, generating a first surface model dataset based on the PDM by modifying at least one weight of the one or more weight-eigenvector pairs, wherein the first surface model dataset is different from the average dataset, and generating an output volume dataset based on the first surface model dataset and a first reference dataset, the first reference dataset comprising data representing the volume of a corresponding part of a body, the output volume dataset comprising data representing a deformed volume of the corresponding part of the body.