Image Analysis System Using Context Features for Biological Samples
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated recognition of objects in biological images faces challenges due to variations in staining intensity and biological heterogeneity, which affect the accuracy of object classification, especially when objects have subtle color differences and varying prevalence across images.
Innovation Solution
An image analysis system that computes context feature values to differentiate between variations caused by object class membership and those due to staining or biological artifacts, using these context features in conjunction with object feature values to improve classification accuracy by leveling out inter-image variability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Stability of the object's composition
If color distribution alignment is applied to improve stain appearance consistency, then stain appearance consistency is improved, but classification accuracy deteriorates due to introduced color confusion between objects
Solution Approach 1:
The patent segments the analysis by computing context features separately for each object class (lymphocytes, tumor cells, non-tumor cells) rather than applying global color alignment. This allows preservation of subtle color differences within each class while normalizing across images, resolving the contradiction between consistency and classification accuracy.
Solution Approach 2:
The patent applies local quality by computing context features specific to each object class and image combination. Instead of uniform global alignment, each object class receives tailored context feature computation that preserves its unique color characteristics while accounting for image-specific variations, thereby maintaining both consistency and classification accuracy.
2Reliability
If global color distribution alignment is performed across images, then inter-image color variation is reduced, but subtle color differences between object classes are lost
Solution Approach 1:
The patent introduces context features as an intermediary that captures image-specific color characteristics without directly modifying the object features. These context features serve as mediators that account for global color shifts while preserving local subtle differences, allowing the classifier to distinguish objects based on relative rather than absolute color values.
3Productivity
If object feature values are used directly for classification without context adjustment, then classification speed is maintained, but classification accuracy deteriorates due to inter-image variability
Solution Approach 1:
The patent applies preliminary action by computing context features for each image before performing classification. This pre-computation captures image-specific variations in advance, allowing the classification step to proceed efficiently using adjusted features that already account for inter-image variability, thus maintaining both speed and accuracy.
Data Source
Figure 1A~1D
Figure 2
Figure 3(a1)~3(b2)
AI summary
The present disclosure relates to an image analysis system for identifying objects belonging to a particular objet class in a digital image (102-108) of a biological sample, the system comprising a processor and memory, the memory comprising interpretable instructions which, when executed by the processor, cause the processor to perform a method comprising: - analyzing (602) the digital image for automatically or semi-automatically identifying objects in the digital image; - analyzing (604) the digital image for identifying, for each object, a first object feature value (202, 702) of a first object feature of said object; - analyzing (606) the digital image for computing one or more first context feature values (204, 704), each first context feature value being a derivative of the first object feature values or of other object feature values of a plurality of the objects in the digital image or being a derivative of a plurality of pixels of the digital image; - inputting (608) both the first object feature value of each of the objects in the digital image and the first context feature value of said digital image into a first classifier (210, 710); and - executing (610) the first classifier, the first classifier thereby using the first object feature value of each object and the one or more first context feature values as input for automatically determining, for said object, a first likelihood (216, 714) of said object of being a member of the object class.