Cognitive Data Processing System for Ambiguously Labeled Medical Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep-learning cognitive systems face challenges in training due to the difficulty in obtaining accurately labeled medical image datasets, often plagued by data entry errors and ambiguities in large electronic medical records, which lead to ambiguously labeled data.
Innovation Solution
The implementation of a cognitive data processing system that utilizes ambiguously labeled data by introducing a novel loss function combining a traditional loss term, a maximizing term, and a sparsity term to model ambiguity, allowing the system to treat ambiguously labeled medical images as definite negative or positive examples, thereby training deep-learning cognitive systems to identify abnormalities or anatomical structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional classification methods are used to train deep-learning cognitive systems, then the system can process clearly labeled data, but it cannot effectively handle ambiguously labeled data from electronic medical records
Solution Approach 1:
The patent changes the parameters of the loss function by introducing three distinct terms (traditional loss term, maximizing term, and sparsity term) that operate on different aspects of the classification problem. This allows the system to handle ambiguous labels by modifying how loss is calculated and optimized during training, enabling effective use of ambiguously labeled medical imaging data while maintaining classification reliability
Solution Approach 2:
The loss function is segmented into three separate terms, each addressing a specific aspect of the classification challenge. The traditional loss term handles standard classification, the maximizing term addresses ambiguous label resolution, and the sparsity term manages label selection. This segmentation allows each term to optimize its specific function while contributing to overall system performance
2Quantity of substance
If ambiguously labeled data is used for training, then more training data becomes available, but the data quality and label accuracy deteriorate
Solution Approach 1:
The patent converts the harmful effect of label ambiguity into a beneficial training signal. By designing the loss function to explicitly model and resolve ambiguities rather than treating them as errors, the system transforms low-quality ambiguously labeled data into valuable training examples that improve the model's ability to handle real-world medical imaging data with inherent uncertainties
3Ease of operation
If a single candidate class is identified for ambiguously labeled data, then classification decision is simplified, but information about multiple possible classes is lost
Solution Approach 1:
The patent implements a dynamic approach where the system first identifies multiple candidate classes that are consistent with the ambiguous labels, then uses the maximizing term to select the most likely single class while preserving information about alternative candidates through the sparsity term. This dynamic process adapts the level of classification specificity based on the ambiguity resolution needs
Data Source
AI summary
A mechanism is provided that implements a cognitive data processing system for automatically processing ambiguously labeled data associated with a medical image. The cognitive data processing system receives an ambiguously labeled set of training data in which the ambiguously labeled set of training data comprises portions of data and associated labels, and wherein at least one portion of data in the ambiguously labeled set of training data has a plurality of different labels that together render the portion of data ambiguously labeled. The cognitive data processing system configures an implementation of a model that comprises a loss term, a maximizing term, and a sparsity term. The cognitive data processing system processes the ambiguously labeled set of training data based on the model to identifying a mapping that minimizes a loss function and thereby train the cognitive data processing system.