Trained Model Creation for Cell Image Segmentation Noise
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing trained models for cell image segmentation often incorrectly label background regions as abnormal cells, leading to decreased detection accuracy, especially when abnormal cells appear less frequently than normal cells.
Innovation Solution
A method and system that utilize machine learning to create a trained model by identifying and labeling noise regions in teacher image data, updating the training data, and re-creating the model to distinguish noise regions from detection targets and backgrounds, thereby improving detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a trained model is created using machine learning to perform segmentation on cell images, then the segmentation process can automatically identify cell regions, but background regions may be erroneously labeled as abnormal cell regions, decreasing detection accuracy
Solution Approach 1:
The patent segments the label image into multiple types including detection target regions, background regions, and noise regions. By dividing the labeling task into distinct categories, the model can more accurately distinguish between actual abnormal cells and background regions that should not be detected, thereby reducing false positives while maintaining automated segmentation.
Solution Approach 2:
The patent changes the parameter of label types from traditional two-class (detection target and background) to three-class (detection target, background, and noise). This parameter change allows the model to specifically identify and separate noise regions from actual detection targets, improving detection accuracy by preventing background regions from being erroneously labeled as abnormal cells.
2Ease of manufacture
If the training data includes only two label types (detection target and background), then the model training is simpler, but the model cannot correctly distinguish noise regions from detection targets, leading to decreased detection accuracy
Solution Approach 1:
The patent modifies the label type parameter from two categories to three categories by introducing noise regions. This allows the model to learn the distinction between actual detection targets and noise regions, significantly improving detection accuracy. The increased complexity is justified by the substantial improvement in distinguishing true abnormal cells from background noise.
Solution Approach 2:
The noise region label acts as an intermediary category between detection targets and background regions. This intermediate classification helps the model understand the nuanced differences between actual abnormal cells and background artifacts, providing a bridge that improves overall detection accuracy without making the training process excessively complex.
3Reliability
If abnormal cells appear less frequently than normal cells, then the dataset is more representative of real-world conditions, but the detection accuracy of abnormal cells decreases due to erroneous labeling of background regions
Solution Approach 1:
By segmenting label regions into detection targets, background, and noise categories, the patent enables the model to specifically identify noise regions that are prevalent in datasets with infrequent abnormal cells. This segmentation prevents background regions from being misclassified as abnormal cells, thereby maintaining detection accuracy even when abnormal cells are rare in the training data.
Solution Approach 2:
The introduction of a third label type (noise region) changes the classification parameter space, allowing the model to better handle the imbalance between frequent background regions and rare abnormal cells. This parameter expansion enables the model to learn from the noise regions without sacrificing the representativeness of the overall dataset.
Data Source
AI summary
A trained model creation method includes creating a trained model (50) based on training data (86) including first teacher image data (86b), determining whether or not an estimated region is erroneously estimated in inference result data (84), determining a noise region (60) included in the first teacher image data (86b), updating the first teacher image data (86b), and re-creating trained model (55).