Automated Nucleated Red Blood Cell Detection via HOG Feature Normalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies face challenges in accurately detecting nucleated red blood cells (NRBCs) in maternal blood due to variations in specimens and imaging conditions, making it difficult to distinguish target cells from non-target cells.
Innovation Solution
An image processing system that includes a device with a preprocessing section for normalization, a feature extraction section using HOG features, and a learning section to determine target cells by comparing image features with learned criteria, reducing the impact of specimen and imaging condition variations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If visual detection method is used to detect NRBCs, then detection capability is maintained, but detection accuracy deteriorates due to heavy strain and difficulty in distinguishing target cells from non-target cells
Solution Approach 1:
The patent replaces the manual visual detection method with an automated image processing system that uses computer algorithms to detect and classify NRBCs. The system captures images of blood cells, processes them through multiple refinement stages including color-based filtering, shape analysis, and machine learning classification, thereby eliminating operator strain while maintaining or improving detection accuracy.
2Measurement precision
If simple image processing is used without normalization and feature extraction, then processing speed is maintained, but detection accuracy deteriorates due to specimen and imaging condition variations
Solution Approach 1:
The patent segments the image processing task into multiple distinct stages: initial image acquisition, color-based candidate selection, shape feature extraction, HOG feature computation, and multi-stage classification. Each stage processes specific features and filters candidates progressively, making the complex processing manageable and effective in distinguishing NRBCs from other cells despite variations in specimens and imaging conditions.
Solution Approach 2:
The patent transforms the input images through various parameter transformations including color space conversion, normalization of image intensity and color values, and extraction of shape and texture features. These parameter changes convert raw pixel data into standardized feature representations that are invariant to imaging condition variations, enabling accurate detection across different specimens and imaging setups.
3Reliability
If no normalization process is applied, then processing time is reduced, but detection reliability deteriorates due to differences in specimens and imaging conditions
Solution Approach 1:
The patent applies normalization and feature extraction processes in advance before the actual classification and detection stages. By pre-processing images to standardize color, intensity, and geometric features, the system prepares the data in a form that is robust to variations in specimens and imaging conditions, thereby improving detection reliability without significantly impacting the time required for the core detection task.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
It is arranged to make it difficult to be affected by the difference between the specimens of the target cell and the difference in imaging conditions and so on. An image processing device 10 obtains a taken image obtained by imaging a sample including the target cell having a nucleus, then extracts pixels to be candidate of the nucleus from pixels included in the taken image based on a first condition determined in advance with respect to at least one of a color and a luminance, which the pixels to be the candidate are required to have, then further extracts a connected pixel group to be a candidate of the target cell from connected pixel groups, each of which is obtained by connecting adjacent ones of the pixels extracted, based on a second condition determined in advance with respect to a size and a shape which the connected pixel group to be the candidate is required to have, then sets a rectangular area, which has a predetermined size centered on the pixel included in the connected pixel group extracted, to the taken image, and then determines whether or not the target cell is included in the rectangular area based on whether or not an image feature amount obtained from the rectangular area thus set satisfies a condition of the image feature amount mechanically learned based on sample images of a positive example and a negative example of the target cell.