Ordinal Classification of Microscope Images via Binary Classifier Scores
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing microscopy systems face challenges in accurately classifying microscope images using binary classifiers, which lose class order information, leading to increased misclassification errors, especially when distinguishing between ordered classes like image quality levels.
Innovation Solution
A computer-implemented method for ordinal classification using a machine-learned model that combines estimates from multiple binary classifiers to form a total score, allowing for classification based on variably definable threshold values and interval limits of different widths, thereby preserving class order and improving precision and flexibility.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple binary classifiers are used to discriminate a larger number of potential classes, then the classification coverage is improved, but the class order information is lost leading to increased misclassification errors
Solution Approach 1:
The patent segments the ordinal classification problem into multiple binary classification sub-problems. Instead of using a single complex classifier, it divides the task into several binary classifiers, each handling a specific threshold comparison. This segmentation allows the system to maintain class order information while achieving comprehensive multi-class discrimination, thereby resolving the contradiction between classification coverage and misclassification error rate.
2Ease of manufacture
If binary classifiers are trained for each class independently, then the training process is simplified, but the class order relationship is ignored reducing classification precision
Solution Approach 1:
The patent introduces dynamic threshold values that can be adjusted based on the ordinal relationship between classes. The binary classifiers are trained with dynamic thresholds that reflect the ordered nature of the classes, allowing the system to maintain training simplicity while improving classification precision through adaptive threshold adjustment that captures class order information.
3Ease of operation
If fixed threshold values are used for classification, then the classification process is straightforward, but the system lacks flexibility for different applications
Solution Approach 1:
The patent performs preliminary action by pre-training the binary classifiers with ordinal relationship information embedded in the training data. The classifiers are pre-adapted to understand the ordered structure of classes through training on labeled data that reflects the ordinal relationships. This preliminary adaptation enables the system to maintain simple operation during inference while achieving high flexibility for different applications through the learned ordinal structure.
Data Source
AI summary
A computer-implemented method for the ordinal classification of at least one microscope image calculates a classification into one of a plurality of classes which form an order with respect to an image property. The microscope image is input into a machine-learned model for ordinal classification. The model comprises binary classifiers which calculate estimates regarding whether the microscope image belongs to cumulative auxiliary classes, which combine different numbers of the classes which follow each other in the order. The estimates can be combined so as to form a total score, wherein the classification occurs by comparing the total score with threshold values which can be defined in a variable manner depending on the application, or interval limits of the classes can be defined so that the classes form intervals of different widths. The model for ordinal classification can also comprise further binary classifiers for inverse auxiliary classes.