Non-uniform Sampling for Unbiased Object Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for estimating structural content in biological tissue, such as cancer cells in tissue slices, are time-consuming and require substantial work, with existing image analysis techniques being biased and unsuitable for biological cell counting.
Innovation Solution
A method involving non-uniform random sampling based on the likelihood of object presence in sectors, where sectors with higher probabilities of object presence are sampled more frequently, allowing for unbiased estimation of structural content using computer image analysis and the Horvitz-Thompson estimator.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If systematic uniform random sampling (SURS) is used to estimate structural content, then unbiased estimation is achieved, but substantial work and time are required
Solution Approach 1:
The patent changes the sampling parameter from uniform probability to non-uniform probability proportional to the estimated structural content in each sector. This allows sectors with higher content to be sampled more frequently, improving estimation precision while reducing the total number of sectors that need to be examined, thus resolving the contradiction between unbiased estimation and processing time
Solution Approach 2:
The patent performs preliminary image analysis on all sectors to estimate the structural content before the actual sampling process. This preliminary action creates a weighting scheme that guides the subsequent sampling, allowing the method to focus resources on sectors most likely to contain objects of interest, thereby reducing overall processing time while maintaining estimation accuracy
2Extent of automation
If existing image analysis techniques are used for object counting, then automation is achieved, but the results are biased and unsuitable for biological cell counting
Solution Approach 1:
The patent introduces an intermediary weighting factor based on estimated structural content that mediates between automated image analysis and final object counting. This intermediary layer corrects the bias inherent in standard image analysis by adjusting the sampling probability according to the likelihood of object presence, thereby maintaining automation while achieving accurate, unbiased counting results
Solution Approach 2:
The patent modifies the sampling parameter from uniform random selection to non-uniform selection proportional to structural content estimates. This parameter change allows automated analysis to focus on high-probability sectors, reducing bias while maintaining the efficiency of automation
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An image is partitioned into sectors, and a number of sectors are selected randomly but with a probability of selection which is proportional with the likelihood of objects in the sector. For the selected sectors, the objects are measured or counted and used for estimation of the amount of objects in the entire image.