GAN-Based Early-Stage Lesion Image Synthesis for Medical AI Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
It is challenging to accurately generate medical images of lesions in their early stages for machine learning, as lesions are often discovered after progression, making it difficult to collect sufficient early-stage data.
Innovation Solution
An image processing apparatus and method that uses a generative adversarial network (GAN) to determine the position of indirect findings associated with lesion occurrence, generating medical images by combining normal and abnormal images, allowing for the creation of early-stage lesion images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If medical images of lesions in early stages are collected for machine learning, then detection accuracy can be improved, but it is difficult to collect sufficient early-stage data because lesions are often discovered after progression
Solution Approach 1:
The patent generates synthetic medical images of lesions in early stages by copying and transforming features from abnormal medical images. The system extracts lesion features from confirmed case images and generates realistic early-stage lesion images through image processing, creating artificial training data that mimics real early-stage pathology without requiring actual early-stage patient data
Solution Approach 2:
The patent performs preliminary generation of early-stage lesion images before they would naturally occur in clinical practice. By generating synthetic early-stage images in advance using machine learning models, the system prepares training data proactively, allowing machine learning models to be trained on early-stage presentations before real-world early-stage cases become available for collection
2Productivity
If a large amount of learning data is collected to improve detection accuracy, then machine learning model performance improves, but the availability of early-stage lesion data remains limited
Solution Approach 1:
The system creates multiple copies of lesion features from limited abnormal images, generating numerous synthetic early-stage images from a small set of source images. This copying approach multiplies the available training data exponentially, allowing robust machine learning model training without requiring extensive collection of rare early-stage clinical cases
Solution Approach 2:
The patent applies parameter changes to transform abnormal medical images into early-stage representations by adjusting features such as lesion size, intensity, and morphological characteristics. By systematically varying these parameters, the system generates diverse synthetic images that represent different early-stage presentations, expanding data availability across multiple scenarios
Data Source
AI summary
An image processing apparatus determines a position of an indirect finding associated with occurrence of a lesion in a medical image based on input information including a position of the lesion, and generates a medical image in which the indirect finding is generated at the determined position in a normal medical image in which the lesion has not occurred or in an abnormal medical image in which the lesion has occurred.