Neural Network Dual-Task Learning for Medical Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition techniques using machine learning struggle with accurately classifying images when the recognition target is not present, leading to incorrect results, as they lack a method to handle images without the target during training, complicating the model and increasing classification classes.
Innovation Solution
An image learning device and method that utilize a neural network with separate tasks for classifying images with and without the recognition target, updating weight coefficients based on specific evaluation results for each task, and employing ground truth data and segmentation labels to prevent incorrect classifications by controlling the loss function and gradient during backpropagation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If images without the recognition target are added to the learning data, then the model can handle negative cases, but the classification model becomes complicated and the number of classification classes increases
Solution Approach 1:
The patent divides the learning process into two separate tasks: a first task that processes images with the recognition target using classification labels, and a second task that processes images without the recognition target using presence/absence labels. This segmentation allows the model to handle negative cases without complicating the classification structure, as each task has its own tailored learning objective and label system.
Solution Approach 2:
The patent introduces a new dimension to the learning framework by adding a second task that operates in parallel to the first task. Instead of expanding the classification classes horizontally, the solution adds a vertical dimension of task differentiation, where the second task specifically addresses images without targets using a simplified presence/absence labeling system, thus avoiding model complexity while improving reliability.
2Device complexity
If the recognizer is trained only on images with the recognition target, then the model is simple, but it outputs incorrect results when images without the target are input
Solution Approach 1:
The patent applies preliminary action by preparing two distinct learning tasks in advance: the first task learns to classify images with targets using classification labels, while the second task learns to identify the presence or absence of targets in images. This preliminary structuring ensures that when images without targets are input, the model has already been trained to handle them appropriately, preventing incorrect results while maintaining relative model simplicity.
Solution Approach 2:
The patent implements feedback mechanisms for both tasks: the first task receives feedback from classification labels to refine its classification accuracy, while the second task receives feedback from presence/absence labels to improve its ability to detect target absence. This dual feedback system ensures that the model learns from both positive and negative cases, improving reliability without significantly increasing complexity.
Data Source
AI summary
An object of the invention is to provide an image learning device, an image learning method, a neural network, and an image classification device which can support appropriate classification of an image.In the image learning device according to an aspect of the invention, the neural network performs a first task of classifying a recognition target in a medical image and outputting a classification score as an evaluation result, and a second task different from the first task. The neural network updates a weight coefficient on the basis of a comparison result between the classification score output for the medical image of a first image group and a ground truth classification label, and does not reflect the classification score output for the medical image of a second image group in an update of the weight coefficient, for the first task. The neural network updates the weight coefficient on the basis of the evaluation result output for the medical image of the first image group and the evaluation result output for the medical image of the second image group, for the second task.