Medical Image Anomaly Training Using Synthetic Abnormal Regions

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing machine learning models for medical image analysis often lack sufficient training data for abnormal conditions, particularly pathologies, making it difficult to detect unseen anomalies effectively.

Innovation Solution

A method and apparatus that generate abnormal medical image data by modifying or replacing local regions within normal medical image data, using techniques like spatial and intensity transformations, to train a model to identify anomalies, even when only normal data is available.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If machine learning models are trained only on normal medical image data, then the model can be trained with widely available data, but the model cannot effectively detect pathologies or abnormal conditions

Engineering Contradiction:

Improveavailability of training dataVSAvoiddetection accuracy for pathologies

Core Design Contradiction:

Quantity of substanceVSReliability

Solution Approach 1:

The system performs preliminary actions by generating synthetic abnormal data samples before actual detection is needed. Normal medical images are pre-processed to create abnormal versions through various transformations, so that when the model needs to detect pathologies, it already has a comprehensive training set that includes both normal and synthetically generated abnormal cases.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system changes parameters of normal medical images to generate abnormal data. By applying spatial transformations (rotation, flipping, cropping), intensity transformations (brightness, contrast adjustments), and anatomical transformations (modifying organ shapes, positions, or characteristics), the system transforms normal images into abnormal ones, enabling the model to learn pathology detection from parameter-modified data.

Inventive Principle:

Principle #35Parameter changes

2Reliability

If abnormal medical image data is generated by modifying normal data, then the model can learn to detect pathologies, but the generated data may not perfectly represent real pathological conditions

Engineering Contradiction:

Improvepathology detection capabilityVSAvoidaccuracy of generated abnormal data

Core Design Contradiction:

ReliabilityVSManufacturing precision

Solution Approach 1:

The system employs dynamic and diverse transformation methods to generate abnormal data. Multiple types of transformations (spatial, intensity, anatomical) are applied with varying parameters and combinations, creating a dynamic training set that captures different manifestations of pathologies rather than static,单一 modifications.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The system creates composite training data by combining multiple transformation techniques. Normal images undergo combinations of spatial transformations, intensity adjustments, and anatomical modifications to generate complex abnormal samples that better approximate real pathological conditions, rather than applying single simple transformations.

Inventive Principle:

Principle #40Composite materials

Data Source

PatentUS12475688B2Data processing apparatus for training a model with abnormal medical image data to identify abnormal medical image data

Publication Date: 2025.11.18 CANON MEDICAL SYST CORP

AI summary

An apparatus for training a model to identify abnormal medical/image data comprises processing circuitry configured to:receive medical/image data;obtain a local region and a context region from the medical/image data;generate abnormal medical/image data using the local region and/or the context region;train a model using the medical/image data and the generated abnormal medical/image data to identify abnormal medical/image data.