Disease recognitionmethod based on lightweight twin convolutional neural network

Abstract

The invention discloses a disease recognition method based on a lightweight twin convolutional neural network. The method comprises the following steps of constructing a fine-grained lesion feature joint training model, wherein the fine-grained lesion feature joint training model comprises a data generator, the data generator is connected with a feature extractor, the feature extractor is connected with the twin convolutional neural network, and the twin convolutional neural network is connected with the feature discrimination network, training the fine-grained lesion feature joint training model, and generating a fine-grained lesion feature recognition model based on the fine-grained lesion feature joint training model with the minimum loss function value, and inputting the to-be-recognized image into the fine-grained lesion feature recognition model, and outputting a corresponding skin disease category. According to the method for carrying out positive and negative sample joint training based on the twin convolutional neural network, the model can extract more discriminative features, the conditions of small inter-class difference and large intra-class difference of lesion imagefeatures in an original data set are effectively relieved, and the feature discrimination capability of the model is enhanced.

Description

technical field [0001] The invention belongs to the technical field of lesion identification, and more specifically, the invention relates to a disease identification method based on a lightweight Siamese convolutional neural network. Background technique [0002] Medical care is related to human life and health. In recent years, the use of data-driven methods to assist medical image analysis and diagnosis has attracted more and more attention from academia and industry in the fields of medical image and computer vision. More and more advanced algorithms have been developed to assist medical image analysis and diagnosis. Human doctors perform disease diagnosis. The automatic recognition model based on machine learning mainly includes two steps of feature extraction and classifier training. Among them, how to effectively model the appearance features directly affects the final performance of the model. In medical image analysis, especially in the field of recognition, the m...

AI Technical Summary

Problems solved by technology

However, it has two main limitations: On the one hand, the lack of medical image standard datasets cannot be alleviated by over-reliance on data enhancement strategies. Doubling amplification, but the effective change mode of new image samples is still relatively limited. At the same time, a large number of repetitive amplification may also bring the risk of overfitting to the model

On the other hand, the target-based detection network usually has a large amount of parameters, and its effective training often requires a certain scale of original data sets as support, which is difficult to implement in medical images. Deployment poses potential difficulties

[0005] b) Model training and feature extraction: The target detection network needs to alternately train different models in the network. The training of the region candidate feature network provides the required candidate regions for further training of the lesion feature location and recognition network. The entire target detection network Effective training requires a good training strategy. At the same time, due to the irregularity and complexity of medical image lesion features, it also brings additional challenges to the selection of candidate frame sizes in the region candidate feature network.

Although the target detection network can realize the positioning and identification of skin lesion areas, its feature extraction, positioning and identification methods lack certain interpretability, and cannot provide effective reasoning basis for lesion feature identification to a certain extent.

Images