Convolutional Neural Network Lesion Labeling via Class-Only Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for determining lesion positions in medical images using traditional feature detection or pattern recognition have low accuracy, and high-accuracy algorithms require labeled images, which are difficult to obtain in practice.
Innovation Solution
A method utilizing a multi-layer convolutional neural network with convolutional, activation, and down-sampling layers to generate a target position mask, where large-value spatial position sampling and mapping are performed to identify target positions in images, allowing for easier training with labeled disease classes rather than specific lesion positions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If traditional feature detection or pattern recognition algorithms are used to determine lesion positions, then the method is simple to implement, but the accuracy is low
Solution Approach 1:
The patent changes the parameter of training data requirements from needing precise lesion position labels to only needing disease class labels. This parameter change allows the use of simpler, more abundant training data while maintaining high accuracy through the neural network's internal feature learning capabilities.
2Measurement precision
If machine learning or deep learning algorithms are used to obtain high-accuracy automatic labeling, then the accuracy is improved, but the training requires labeled images with lesion positions which are difficult to acquire
Solution Approach 1:
The patent extracts and removes the requirement for precise lesion position labels from the training process. By using only disease class labels and modifying the neural network architecture to perform both classification and localization, the patent eliminates the bottleneck of acquiring difficult-to-obtain labeled data while maintaining high accuracy.
3Measurement precision
If a neural network is trained with labeled lesion positions, then high-accuracy target labeling can be achieved, but the data acquisition and labeling process becomes complex and time-consuming
Solution Approach 1:
The patent performs preliminary action by training the neural network to recognize disease classes first, without requiring precise lesion position annotations. The network learns disease characteristics and patterns in advance, and then uses this knowledge to automatically locate targets, thereby avoiding the time-consuming process of manual lesion position labeling.
Data Source
AI summary
The present disclosure provides a method and apparatus of labeling a target in an image, and a computer recording medium. The method includes: acquiring a first neural network, the first neural network includes a multi-layer convolutional neural network and a fully connected layer, wherein each layer of the multi-layer convolutional neural network includes a convolutional layer, an activation function layer and a down-sampling layer arranged successively; processing the image by using the multi-layer convolutional neural network of the first neural network acquired so as to obtain a target position mask for the image; and labeling the target in the image based on the target position mask.