Dual-Class Neural Network for Disease Diagnosis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional disease diagnosis systems using neural networks are limited in accuracy as they are trained to recognize only one disease state per unitary unit, neglecting the possibility of multiple states, which can lead to ambiguous classifications and reduced performance.
Innovation Solution
A disease diagnosis system that incorporates a loss function allowing for dual labeling of unitary units, enabling the neural network to output probabilities for multiple states, with a specific layer calculating losses based on feature values for both primary and secondary states, improving the accuracy of disease detection at the pixel level.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If the neural network is trained to determine only one disease state per unitary unit, then the training process is simplified, but the accuracy of disease diagnosis is reduced due to neglecting the possibility of multiple states
Solution Approach 1:
The patent changes the parameter of state representation from single-state classification to multi-state probability distribution. Each unitary unit is represented by multiple feature values corresponding to different disease states, allowing the system to capture ambiguous cases where a unit may belong to multiple states simultaneously. This parameter change resolves the contradiction by enabling accurate representation of complex pathological conditions without oversimplifying the training process.
Solution Approach 2:
The patent adds a dimensional aspect to the classification problem by introducing multiple disease state dimensions for each unitary unit. Instead of forcing a single-class classification, the system outputs a vector of feature values representing probabilities or scores for multiple disease states. This dimensional expansion allows the neural network to preserve information about ambiguous cases while maintaining a structured training framework.
2Ease of operation
If the neural network outputs a single disease state for each unitary unit, then the output interpretation is straightforward, but the ability to detect ambiguous cases with multiple possible states is lost
Solution Approach 1:
The patent introduces dynamics into the output representation by allowing the system to adaptively determine the number and nature of disease states for each unitary unit based on the input characteristics. The neural network can dynamically adjust the feature values to reflect the degree of ambiguity or confidence in each disease state classification, enabling reliable detection of complex cases while maintaining interpretability through the structured output format.
Solution Approach 2:
The patent uses feature values as intermediaries between the neural network's internal processing and the final disease state interpretation. These feature values serve as a bridge that preserves detailed information about multiple possible states while allowing for systematic interpretation. The feature values can be processed further to generate clinically meaningful outputs, maintaining both reliability and ease of operation.
3Ease of manufacture
If conventional single-state labeling is used for training data, then the data preparation process is simpler, but the neural network cannot learn from ambiguous cases with multiple possible states
Solution Approach 1:
The patent changes the parameter of data representation from single-label to multi-label format. Each training example is represented by multiple feature values corresponding to different disease states, allowing the neural network to learn from the nuanced relationships between input images and multiple possible diagnoses. This parameter change enables the system to utilize ambiguous cases as training data, improving learning effectiveness without significantly complicating the data preparation process.
Solution Approach 2:
The patent creates a universal training data format that can represent both clear-cut single-state cases and ambiguous multi-state cases within the same framework. The multi-state feature value representation serves multiple functions: it captures definitive diagnoses, represents ambiguous cases, and provides gradient information for learning. This universal format enables the neural network to learn from diverse training examples uniformly, improving overall learning effectiveness.
Data Source
AI summary
A disease diagnosis system includes a processor and a storage device storing a neural network. The processor trains the neural network in the storage device to output a determination value corresponding to a probability having at least one of a plurality of states using a given loss function and learning data labeled so that a given unitary unit included in a biometric image is to have at least one of the plurality of states. The neural network includes a specific layer to output a plurality of feature values corresponding to a probability that the unitary unit is to be determined as each of the plurality of states. The loss function incorporates both first and second feature values corresponding to first and second states into a dual labeling unitary unit with the first state having a higher probability and a second state having lower probability.