Synthetic Patient Model Generation via Anatomical Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for training practitioners and machine learning models to analyze medical images face limitations due to small and unrepresentative data sets, leading to inadequate accuracy and reliability in medical image analysis.
Innovation Solution
A computer-implemented method for creating a synthetic patient model using a generative model, which involves obtaining medical information, generating an anatomical map, conditioning the generative model, embedding medical information, and creating synthetic patient images that mimic real patient images with specified anatomical structures and imaging characteristics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If real patient data sets are used for training, then training accuracy may be improved, but data availability and representativeness deteriorate due to small sample sizes and difficulty in retrieving pertinent data
Solution Approach 1:
The patent creates synthetic copies of patient data through generative models. The system generates artificial patient records, images, and clinical data that replicate the statistical properties and relationships of real patient data without using actual patient information. This copying approach solves the contradiction by providing unlimited training data (improving quantity) while maintaining the realistic patterns needed for accurate training (preserving measurement precision).
Solution Approach 2:
The patent transforms the nature of training data by changing from real patient data to synthetically generated data. The generative model adjusts various parameters including patient demographics, clinical measurements, imaging characteristics, and outcome variables to create realistic synthetic datasets. This parameter transformation allows unlimited data generation (improving quantity) while maintaining statistical validity (preserving training accuracy).
2Adaptability or versatility
If more real patient data is collected to improve representativeness, then data representativeness improves, but patient privacy and data security deteriorate
Solution Approach 1:
The patent creates synthetic copies of patient data that preserve statistical representativeness without containing actual patient information. The generative model learns the distribution and relationships in real data, then generates artificial samples that match these patterns. This approach achieves full data representativeness (improving adaptability) while completely eliminating patient privacy risks (removing harmful factors) since no real patient data is stored or processed after the initial training phase.
Solution Approach 2:
The generative model acts as an intermediary between real patient data and training applications. It receives real data only during the initial training phase, then generates synthetic data for all subsequent use. This intermediary layer preserves the statistical properties needed for representativeness while blocking direct access to real patient information, thereby eliminating privacy risks in deployment scenarios.
3Productivity
If synthetic data is generated using simple methods, then generation speed improves, but data quality and realism deteriorate
Solution Approach 1:
The patent implements continuous synthetic data generation through the trained generative model. Once trained on real data, the model can indefinitely generate high-quality synthetic patient records without degradation. The system maintains continuous operation by generating data on-demand with consistent quality and realism, solving the contradiction by achieving both high generation speed (improving productivity) and high data quality (preserving measurement precision) through the efficiency of the trained generative process.
Solution Approach 2:
The patent performs preliminary training of the generative model on real patient data before deployment. This preliminary action captures the complex relationships and distributions in real data, enabling the model to generate high-quality synthetic data efficiently thereafter. The upfront training investment (preliminary action) creates a model that subsequently produces both fast generation (improving productivity) and realistic output (preserving data quality) without requiring complex real-time processing.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
Provided is a computer-implemented method for creating a synthetic patient model comprising a synthetic patient image. The method comprises: obtaining a data set comprising medical information, wherein the medical information comprises an anatomical description comprising at least information specifying an anatomical structure to be depicted in the synthetic patient image, and imaging information specifying image formation characteristics of the synthetic patient image, and wherein the medical information optionally comprises anamnesis and physical examination descriptions; generating an anatomical map of the anatomical structure by means of anatomy modelling using the anatomical description as an input; conditioning a generative model using the anatomical map; embedding a subset of the medical information, the subset comprising textual information and including at least the imaging information, in the generative model; creating the synthetic patient model including a synthetic patient image by means of the generative model.