Medical Image Registration for Anatomically Faithful Data Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical image data augmentation methods fail to effectively increase variation in learning images, leading to reduced prediction accuracy and reliability in disease opinions due to blurring of anatomical structures, particularly in medical images.
Innovation Solution
Perform non-linear registration processing on pairs of medical images to generate new images by transforming them based on transformation coefficients, maintaining anatomical structure integrity and varying the classes of the images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data augmentation methods such as parallel displacement, rotation, enlargement/reduction, inversion, cropping, and noise addition are applied to generate new learning images, then the number of learning images is increased, but the variation of learning images is not sufficiently increased and anatomical structures become blurred
Solution Approach 1:
The patent applies non-linear registration processing with transformation parameters (transformation amount T_12, T 21 and transformation coefficients α, β) to generate new medical images. By changing these parameters systematically, the patent generates diverse learning images while preserving anatomical structures, resolving the contradiction between increasing image quantity and maintaining reliability.
Solution Approach 2:
The patent replaces traditional mechanical image processing methods (parallel displacement, rotation, etc.) with non-linear registration processing based on mathematical transformation functions. This substitution enables more sophisticated image transformation that maintains anatomical integrity while generating sufficient variation in learning images.
2Quantity of substance
If traditional data augmentation methods are used to increase the number of learning images, then the learning dataset size is expanded, but the anatomical structures in the generated images become blurred
Solution Approach 1:
The patent uses non-linear registration processing with controllable transformation parameters to generate new medical images. The transformation amount T and coefficients α, β are carefully controlled to ensure that anatomical structures remain clear and recognizable, thus expanding the learning dataset without sacrificing anatomical precision.
Solution Approach 2:
The patent introduces non-linear registration processing as an intermediary method between original medical images and generated learning images. This intermediary process ensures that anatomical structures are preserved through mathematically controlled transformations, unlike direct mechanical processing methods that cause blurring.
3Adaptability or versatility
If non-linear registration processing is performed with large transformation amounts to increase image variation, then more diverse learning images are generated, but the anatomical structures become distorted
Solution Approach 1:
The patent systematically adjusts transformation parameters (transformation amount T and coefficients α, β) to achieve optimal balance. By changing these parameters within controlled ranges, the patent generates diverse learning images while maintaining anatomical structure integrity, resolving the contradiction between variation and shape preservation.
Solution Approach 2:
The patent applies dynamic transformation coefficients (α, β) that can be adjusted based on the specific images being processed. This dynamic approach allows the system to adapt the transformation strength to preserve anatomical structures while still generating sufficient variation for effective machine learning training.
Data Source
AI summary
An image processing apparatus includes a processor and a memory connected to or built in the processor. The processor is configured to perform non-linear registration processing on a first medical image and a second medical image among a plurality of medical images, and generate at least one new medical image that is used for training a machine learning model for the medical images by transforming at least one medical image of the first medical image or the second medical image based on a result of the non-linear registration processing.