Pseudo Image Generation for Rare Object Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models face challenges in accurately segmenting rare objects in medical images due to the scarcity of training data, particularly for progressive diseases like cancer, where cancer tissue often infiltrates surrounding areas, making it difficult to generate sufficient training data for accurate segmentation and classification.
Innovation Solution
An image generation apparatus that processes mask images to derive pseudo mask images and pseudo images, which are then used as training data to construct a segmentation model capable of segmenting rare objects by generating images with varying shapes and progressions of lesions, including those not present in existing data, thereby enriching the training dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing machine learning models are trained only on real medical images, then the model can accurately segment common objects, but it cannot accurately segment rare objects or progressive diseases due to insufficient training data
Solution Approach 1:
The patent creates pseudo-images by copying and modifying existing medical images. Specifically, it generates synthetic images of rare diseases and progressive cancer stages by manipulating pixel data and mask information from real images, thereby expanding the training dataset without requiring actual rare disease samples
Solution Approach 2:
The patent changes parameters of existing images to create new training data. It modifies image characteristics such as lesion progression stages, tumor size, and spatial distribution by adjusting mask values and pixel intensities, transforming common disease images into pseudo-images representing rare conditions
2Quantity of substance
If pseudo-images are generated from existing images, then training data quantity increases, but the ability to generate data for features rarely included in existing learning data is limited
Solution Approach 1:
The patent segments the image processing into distinct components: it separates the mask image into multiple mask images corresponding to different disease stages or characteristics, then processes each mask independently to generate diverse pseudo-images. This segmentation enables systematic exploration of different disease features
Solution Approach 2:
The patent adds new dimensions to the training data by generating pseudo-images that represent disease progression stages, tumor heterogeneity, and spatial variations not present in the original images. It transforms 2D medical images into multi-dimensional training data representing different disease states
3Adaptability or versatility
If mask processing is applied to generate pseudo mask images, then diverse training data can be created, but the complexity of the image generation process increases
Solution Approach 1:
The patent performs preliminary processing by pre-processing the mask image to create a pseudo mask image before generating the final pseudo-image. This preliminary action involves thresholding, morphological operations, and mask refinement that simplifies subsequent image generation steps
Solution Approach 2:
The patent uses the pseudo mask image as an intermediary between the original mask and the final pseudo-image. The pseudo mask serves as a mediator that guides the image generation process, allowing complex disease representations to be created through systematic mask manipulation rather than direct complex processing
Data Source
AI summary
A processor is configured to acquire an original image and a mask image in which masks are applied to one or more regions respectively representing one or more objects including a target object in the original image, derive a pseudo mask image by processing the mask in the mask image, and derive a pseudo image that has a region based on a mask included in the pseudo mask image and has the same representation format as the original image, based on the original image and the pseudo mask image.