Medical Image Recognition Model Using Frozen Convolution Layers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The limited availability of medical images in public databases hinders the improvement of pre-trained models for medical image recognition, leading to reduced accuracy in classification, target detection, and segmentation tasks.
Innovation Solution
A method and system are developed to establish pre-trained models using a large number of medical images from specific domains like X-ray, CT, MRI, and pathological photography, training models in a multi-task manner with disease markers to enhance recognition accuracy, where parameters of earlier convolution layers are frozen to maintain feature similarity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If pre-trained models are established using general-purpose databases like ImageNet, then the model structure can be built with large number of images, but the accuracy of medical image recognition is reduced due to lack of medical-specific training data
Solution Approach 1:
The patent applies preliminary action by first training a pre-trained model on general-purpose ImageNet database to establish a base structure with learned features, then subsequently fine-tuning it on medical images. This two-stage approach allows the model to first acquire general image recognition capabilities and then specialize for medical applications, resolving the contradiction between having sufficient training data and achieving domain-specific accuracy
Solution Approach 2:
The patent implements local quality by freezing the parameters of early convolution layers (which learn general features) while allowing later layers to be trained on medical-specific data. This selective training approach preserves the general features learned from large-scale ImageNet while adapting the model to medical image characteristics, thus maintaining both the benefit of large training data and achieving medical recognition accuracy
2Adaptability or versatility
If pre-trained models are established using limited medical images, then the model can be specialized for medical applications, but the accuracy cannot be improved due to insufficient training data
Solution Approach 1:
The patent merges two data sources - general-purpose ImageNet database and medical image database - into a unified training process. By combining the strengths of both datasets through a two-stage training approach (first on ImageNet, then on medical images), the model achieves both adaptability to medical domains and high recognition accuracy, overcoming the limitation of insufficient medical training data
Solution Approach 2:
The patent uses preliminary action by first establishing a pre-trained model on large-scale general images to create a robust base structure, then applying medical-specific fine-tuning. This preliminary training on abundant general data provides a strong foundation that can be subsequently adapted to medical applications, resolving the contradiction between domain specialization and sufficient training data
3Adaptability or versatility
If all layers of the neural network are trained on medical images, then the model can be fully adapted to medical applications, but the training time and computational resources increase significantly
Solution Approach 1:
The patent applies local quality by selectively training only certain layers of the neural network on medical images while freezing others. Specifically, early convolution layers are frozen to preserve general features, while later layers are trained to capture medical-specific patterns. This selective approach achieves adequate medical domain adaptation while significantly reducing training time and computational resources compared to full-network training
4Measurement precision
If early convolution layers are frozen during training, then the general features learned from ImageNet are preserved, but the model's ability to learn medical-specific features may be limited
Solution Approach 1:
The patent implements local quality by applying different training strategies to different parts of the network: early convolution layers are frozen to preserve general features, while later layers are trained on medical images to learn domain-specific features. This spatial differentiation in training approach ensures both feature similarity (through freezing) and medical adaptability (through selective training of later layers)
Solution Approach 2:
The patent segments the neural network into different functional parts with different training regimes. Early layers that capture general features are frozen, while later layers that need to capture medical-specific patterns are trained. This segmentation allows the model to maintain both general feature extraction capabilities and medical domain adaptability simultaneously
Data Source
AI summary
The present disclosure provides a method and a system for recognizing medical image, the present disclosure utilizes the image data with markers of different diseases for calculating and analyzing to build a pre-trained model, the present disclosure has significant improvements to improve the accuracy of image recognition under the general situation of insufficient effective data in the field of medical image recognition technology, the present disclosure can be applied to the field of medical image recognition technology, including X-ray, CT, MRI, ultrasonic, pathological slice photography or fundus photography.