Hierarchical Label Architecture for Biological Tissue Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current molecular and genomic profiling methods for cancer detection are hindered by long waiting periods, high costs, and the need for large tissue samples, while deep learning methods face challenges in accurately classifying biological images due to complex morphological signatures, leading to unreliable predictions.
Innovation Solution
A hierarchical label architecture is introduced, where images of biological tissues are classified using multiple infra-level label classification models that define sub-characteristics, allowing for the determination of upper-level biological characteristics without directly training on them, leveraging the biological relation between characteristics to improve accuracy and reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep learning methods are applied to classify biological images according to upper-label biological characteristics, then prediction capability is provided, but classification accuracy is insufficient due to complex and variable morphological signatures
Solution Approach 1:
The patent segments the classification task by introducing a hierarchical label architecture that divides upper-level biological characteristics into multiple infra-level characteristics. Instead of directly classifying complex upper-label characteristics, the system first classifies simpler infra-level characteristics (e.g., cellular morphology, tissue architecture) and then combines these classifications to determine upper-level characteristics. This segmentation reduces the complexity of the classification problem at each level, thereby improving classification accuracy while maintaining prediction capability.
Solution Approach 2:
The patent adds a hierarchical dimension to the classification system by introducing multiple levels of labels (infra-level and upper-level). This dimensional transformation allows the system to approach the classification problem from different levels of abstraction, where infra-level classifications provide foundational information that supports more accurate upper-level predictions, thus resolving the accuracy-versus-versatility contradiction.
2Measurement precision
If molecular and genomic profiling methods are used to detect cancer, then accurate molecular characteristics are obtained, but long waiting periods reduce survival chances
Solution Approach 1:
The patent replaces traditional molecular and genomic profiling methods (which require laboratory processing and sequencing) with a computational image analysis system using deep learning. By substituting physical molecular analysis with computational classification of histopathology images, the system maintains accurate molecular characteristic detection while eliminating the time-consuming laboratory processes, thus reducing waiting periods and improving patient outcomes.
3Ease of operation
If deep learning models are trained directly on upper-label biological characteristics, then classification is attempted, but reliability is insufficient due to complex morphological signatures
Solution Approach 1:
The patent applies preliminary action by performing infra-level classifications before attempting upper-level classification. The system first trains and executes classification models for infra-level characteristics, stores these results, and then uses them as input for upper-level classification. This preliminary classification step prepares the data in a more reliable format, improving the overall reliability of the classification process while maintaining operational simplicity through automated hierarchical processing.
Data Source
Figure 1~2
AI summary
A method to classify an image of a biological tissue of a patient according to a label relative to a biological characteristic of the biological tissue, this label, relative to the biological characteristic of the biological tissue, being an upper-label according to a label hierarchical architecture, the hierarchical label architecture comprising : - a first lower hierarchical label level comprising : a plurality of first infra-level labels, each being relative to a first lower level biological characteristic of the biological tissue, - a upper hierarchical label level comprising : an upper-level label, relative to the biological characteristic of the biological tissue , defined by a combination of the plurality of infra-level labels, the biological characteristics of the biological tissue being determined by the first lower level biological characteristics of the biological tissue; the method comprising: - a classification module implements a plurality of infra-level label classification models on the image of biological tissue, each infra-label classification model being relative to a first infra-level label of the first lower hierarchical label level relative to a first lower level biological characteristic of the biological tissue; the classification output of each of the infra-level label classification models implemented being relative to a first lower level biological characteristic of the biological tissue.