NRBC Discriminator Training via User-Validated Image Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in diagnosing prenatal fetuses lies in the burdensome visual detection of infinitesimal fetus-derived nucleated red blood cells (NRBCs) in maternal blood, requiring efficient image processing techniques to accurately distinguish target cells from non-target cells.
Innovation Solution
An image processing system that includes an optical microscope, image processing apparatus, and display apparatus, which captures and processes images of maternal blood, extracts nucleus candidate areas, defines cell candidate areas, calculates image feature values, and trains a discriminator using user-inspected samples to improve discrimination accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If visual detection of NRBCs is performed manually, then detection accuracy can be maintained, but the process becomes burdensome and time-consuming
Solution Approach 1:
The patent replaces manual visual detection with an automated image processing system that uses computer algorithms to identify and classify NRBCs. The system captures images of maternal blood, processes them through multiple filtering stages (color segmentation, shape analysis, texture evaluation), and automatically determines the presence of NRBCs, thereby eliminating the time burden of manual inspection while maintaining detection accuracy.
2Productivity
If mechanical detection methods are used to search for NRBCs based on color, shape, and area ratio, then detection efficiency is improved, but discrimination accuracy between target and non-target cells may be insufficient
Solution Approach 1:
The patent divides the detection process into multiple sequential stages: initial color-based segmentation to identify candidate cells, shape analysis to filter candidates, texture evaluation to further refine selections, and final classification by the discriminator. This multi-stage segmentation approach allows the system to efficiently process large numbers of cells while progressively improving discrimination accuracy at each stage.
Solution Approach 2:
The system incorporates a discriminator that is trained using feedback from user inspections. When users verify or correct the system's classifications, this feedback is used to retrain and improve the discriminator's accuracy over time. This feedback mechanism allows the system to maintain high detection efficiency while continuously improving its discrimination capability between target and non-target cells.
3Reliability
If training samples are manually specified for discriminator training, then training can be performed, but time and effort are required for specifying training samples
Solution Approach 1:
The system enables users to train the discriminator by simply inspecting and validating cell images without requiring manual specification of training samples. The system automatically captures images, processes them through the detection pipeline, and presents processed results for user validation. Users provide feedback by confirming correct classifications or correcting misclassifications, and the system automatically uses this feedback to retrain the discriminator, eliminating the time-consuming task of manually selecting and labeling training samples.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
To train a discriminator which discriminates presence of a target cell from absence of a target cell in an image, in a process in which a user inspects cells, an image-feature-value calculating unit 18 extracts an image feature value of the image of a cell candidate area. An NRBC discriminating unit 20 uses the pre-trained discriminator to identify whether or not a target cell is shown in the cell candidate area, on the basis of the image feature value of the image of the cell candidate area. When the cell candidate area is identified as an area in which a target cell is shown, a discrimination result display unit 26 displays the image of the cell candidate area. When the cell candidate area is identified as an area in which a target cell is shown, a discriminator training unit 34 trains the discriminator by using the image feature value of the image of the cell candidate area as a training sample on the basis of a user input about whether or not a target cell is shown in the cell candidate area.