AI Particle Classification via Proportion-Based Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The challenge in microbiology is the tedious process of obtaining annotated training images for supervised learning neural networks, which requires manual annotation of biological particles, making it time-consuming and labor-intensive.

Innovation Solution

A method that uses an artificial intelligence algorithm trained by proportions of labels, where particles are annotated based on their proportions within a class, reducing the need for individual annotations and allowing for adaptive classification as new particles are acquired.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If manual individual annotation of particles is performed to train supervised learning neural networks, then the neural network can be properly trained to identify and classify particles, but the process becomes extremely time-consuming and labor-intensive

Engineering Contradiction:
Improvetraining data qualityVSAvoidannotation time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The system uses automated algorithms to perform annotation based on proportion labels rather than requiring manual individual particle annotation. The neural network learns from proportion-based annotations of entire images, allowing the system to self-train without extensive human intervention for each particle

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The annotation approach changes from individual particle-level labels to image-level proportion labels. Instead of annotating each particle individually, the system uses proportion annotations (e.g., 70% positive, 30% negative) for entire images, fundamentally changing the parameter structure of training data

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If exhaustive individual annotations are obtained for all particles, then the neural network achieves high classification accuracy, but the complexity and cost of data preparation increases significantly

Engineering Contradiction:
Improveclassification accuracyVSAvoiddata preparation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The method extracts only the essential information needed for training by using proportion labels at the image level rather than requiring detailed individual particle annotations. This extraction approach captures the necessary training signal while avoiding the complexity of exhaustive annotation

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

Instead of performing complete individual annotation for every particle, the system uses partial annotation at the image level with proportion labels. This partial action approach provides sufficient training information without the excessive effort of full individual particle annotation

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4394730A1Method for processing an image of a sample comprising biological particles
Publication Date: 2024.07.03 COMMISSARIAT A LENERGIE ATOMIQUE ET AUX ENERGIES ALTERNATIVES
  • EP4394730A1 patent drawingFigure 1A
  • EP4394730A1 patent drawingFigure 1B
  • EP4394730A1 patent drawingFigure 2A~2C

AI summary

A method for characterizing biological particles from a sample, the characterization aimed at determining a property of each particle, the method comprising: - a) measuring a characteristic of at least one particle of the sample; - b) processing the characteristic of the particle or particles by an artificial intelligence algorithm; - c) from the processing, characterizing the particles of the sample, so as to assign a class to each particle, each class being representative of the property of the particle; - the method being characterized in that the artificial intelligence algorithm has previously been trained by labels, from training samples, to each training sample being assigned a proportion of cells in each class, so that the learning is carried out according to the proportions respectively assigned to each training sample.