Learning Image Generation Device for Brain Disease Segmentation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

In end-to-end deep learning for image segmentation, particularly with brain images showing diseases like cerebral apoplexy, it is challenging to prepare learning images that cover the variety of pixel values and definitions of diseased regions due to their undefined shape, size, and location, as well as varying pixel values over time.

Innovation Solution

A learning image generation device and method that acquire supervised data including a learning image and a correct learning image, and generate variation learning images by varying the pixel values of the correct region within a limitation range, allowing for the creation of diverse pixel value representations from a limited learning image.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If a limited learning image is used, then the amount of data required is reduced, but the ability to cover various pixel values of the correct region is insufficient

Engineering Contradiction:

Improveamount of learning dataVSAvoidcoverage of pixel values

Core Design Contradiction:

Quantity of substanceVSAdaptability or versatility

Solution Approach 1:

The patent creates synthetic copies of the limited learning image by generating variation learning images. These copies have the same basic structure and correct region definition as the original learning image, but with varied pixel values within the allowable range. This allows the model to learn from multiple diverse examples without requiring multiple original images, thus resolving the contradiction between limited data quantity and need for diverse pixel value coverage.

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The patent applies parameter changes by varying the pixel values of pixels belonging to the correct region within a predetermined allowable range. By systematically changing these parameters (pixel values) while maintaining the correct region definition, the system generates diverse learning examples from a single image. This transforms the limited learning image into multiple variation learning images with different pixel characteristics, effectively increasing the coverage of pixel values without increasing the number of original images.

Inventive Principle:

Principle #35Parameter changes

2Measurement precision

If end-to-end deep learning is used for semantic segmentation, then classification accuracy is improved, but the difficulty of preparing learning images with defined correct regions increases

Engineering Contradiction:

Improveclassification accuracyVSAvoidpreparation of learning images

Core Design Contradiction:

Measurement precisionVSEase of manufacture

Solution Approach 1:

The patent applies preliminary action by pre-defining the correct region in the learning image before the deep learning process. Instead of requiring the model to learn segmentation from raw images without guidance, the correct region is manually or automatically defined in advance as part of the supervised data preparation. This preliminary definition of the correct region simplifies the learning process and improves classification accuracy while making the preparation of learning images more straightforward.

Inventive Principle:

Principle #10Preliminary action

3Adaptability or versatility

If brain images with undefined disease regions are used, then the applicability to real medical data is improved, but the ability to define correct regions for learning is reduced

Engineering Contradiction:

Improveapplicability to real medical dataVSAvoiddefinition of correct regions

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent introduces an intermediary approach by using variation learning images as a bridge between real medical data and learning requirements. The variation learning images are generated from real brain images by systematically varying pixel values within allowable ranges, creating synthetic data that maintains the characteristics of real medical images while providing well-defined correct regions. This intermediary synthetic data allows the model to learn from real-world applicability while having clearly defined correct regions for supervised learning.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS12217488B2Learning image generation device, learning image generation method, learning image generation program, learning method, learning device, and learning program

Publication Date: 2025.02.04 FUJIFILM CORP

AI summary

A learning image generation device including: a supervised data acquisition unit that acquires supervised data including a learning image and a correct learning image in which a correct region is defined in the learning image as a pair; and a variation learning image generation unit that generates a variation learning image in which a pixel value of a pixel belonging to the correct region is varied within a limitation range of an allowable pixel value of the pixel belonging to the correct region in the learning image.