GAN Synthetic Data for Tumor Lesion Characterization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for tumor lesion characterization, particularly in lung nodules and colorectal cancer, face challenges due to limited data availability, relying heavily on radiologist experience and manual feature extraction, which are inefficient and do not allow for real-time adaptation of treatment based on tumor growth predictions.

Innovation Solution

The use of Generative Adversarial Networks (GANs) to generate synthetic training data for deep learning algorithms, combined with Adaptive Sample Weighting (ASW) and sparse representations, enables effective tumor lesion characterization with limited data, allowing for improved classification performance and early assessment of tumor growth.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If deep learning algorithms are used for tumor lesion characterization, then classification performance is improved, but data availability becomes insufficient

Engineering Contradiction:
Improveclassification performanceVSAvoiddata availability
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent uses Generative Adversarial Networks (GANs) to generate synthetic copies of medical images that mimic real patient data. The generator network creates artificial images while the discriminator network evaluates their authenticity, producing synthetic training data that expands the available dataset without requiring additional real patient scans.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent applies Adaptive Sample Weighting (ASW) to dynamically adjust the importance weights of different training samples during the learning process. This allows the model to adaptively focus on more informative or difficult samples, optimizing training efficiency and performance with limited data by changing the parameter weights rather than requiring more data.

Inventive Principle:
Principle #35Parameter changes

2Quantity of substance

If manual feature extraction by radiologists is used, then data requirements are reduced, but classification efficiency and accuracy deteriorate

Engineering Contradiction:
Improvedata requirementsVSAvoidclassification efficiency
Core Design Contradiction:
Quantity of substanceVSProductivity

Solution Approach 1:

The patent replaces the manual mechanical process of radiologist feature extraction with an automated deep learning system. Instead of radiologists manually analyzing images and extracting features, the system uses trained neural networks to automatically perform feature extraction and classification, significantly improving efficiency while maintaining or enhancing accuracy.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The deep learning system performs self-service by automatically extracting features and making classifications without requiring manual radiologist intervention for each case. The model learns from training data and then independently characterizes new tumor lesions, reducing the time and labor required while improving consistency.

Inventive Principle:
Principle #25Self-service

3Measurement precision

If extensive training data is collected, then model performance is improved, but training time and computational resources increase

Engineering Contradiction:
Improvemodel performanceVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent uses Adaptive Sample Weighting to dynamically adjust training parameters during the learning process. By changing the weights assigned to different samples based on their informational value and difficulty, the system optimizes the training process to achieve better performance with fewer iterations and less computational time than traditional uniform weighting approaches.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent generates only the necessary amount of synthetic data through GANs rather than collecting or generating exhaustive datasets. The ASW technique also focuses computational effort on the most informative samples, applying partial action principles to achieve sufficient model performance without the excessive time and resources required for comprehensive data collection and uniform training approaches.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3576020A1Methods for generating synthetic training data and for training deep learning algorithms for tumor lesion characterization, method and system for tumor lesion characterization, computer program and electronically readable storage medium
Publication Date: 2019.12.04 SIEMENS HEALTHINEERS AG
  • EP3576020A1 patent drawingFigure 1
  • EP3576020A1 patent drawingFigure 2
  • EP3576020A1 patent drawingFigure 3

AI summary

Method for generating synthetic training data for training a deep learning algorithm (1), comprising the steps of: - training a Generative Adversarial Network (5) to generate synthetic image data (4), wherein the Generative Adversarial Network (5) comprises a generator (6) and a discriminator (7), - using the generator (6) of the Generative Adversarial Network (5) to generate synthetic image data (4) as the synthetic training data.