Thyroid Tumor Recognition And Image Segmentation Based On Machine Learnin

Thyroid cancer is a common malignant tumor,but there is still a certain rate of misdiagnosis in clinic.The key to reducing the rate of misdiagnosis lies in improving the accuracy of image analysis.Computer-aided diagnosis methods based on machine learning have greatly improved the accuracy of medical image analysis in recent years.However,medical image data with highquality manual annotation is very scarce,and medical images have huge differences from natural images in feature representation and color space.Therefore,how to learn effective feature representations from a limited data set has become the key to applying machine learning techniques to medical image analysis.In addition,the current research on computer-aided diagnosis of thyroid cancer mainly focuses on ultrasound images,and there is no extensive research on CT images.In order to cope with these problems,this paper carries out the computeraided diagnosis research of thyroid cancer based on CT images.The main research content includes thyroid nodule classification and thyroid image segmentation,where image segmentation is the basis of nodule classification.The specific research contents are:(1)classification of thyroid nodules based on texture analysis;(2)classification of thyroid nodules based on deep pretrained model;(3)thyroid image segmentation based on convolutional recurrent neural network.In particular,this paper innovatively proposes an image preprocessing method based on pseudocolor mapping,which greatly improves the efficiency of feature extraction of the deep pre-trained model on CT images.The main work of this article is as follows:1)To solve the problem of a large number of redundant and irrelevant features extracted by the texture analysis method,an embedded feature selection method based on random forest was used,and the effective features are selected according to the importance of feature based on Gini impurity.After applying this feature selection method,the number of features decreased by 65%.Experiments on the classification of benign and malignant thyroid nodules in CT images through various machine learning classification algorithms show that this method can improve the classification performance of most of the algorithms in the experiment to a certain extent,among them the classification accuracy of the support vector machine algorithm was improved from 92.67 % to 95.26%.2)To solve the problem of domain adaptability of the deep pre-trained model in medical image analysis and the difference in color space between medical images and natural images,a VGG-GAP network model based on the deep pre-trained model and a Pseudo-color mapping grayscale medical image preprocessing method were proposed.Through experiments,it was found that compared with Tiny Net,a small convolutional model trained from scratch,VGG-GAP based on pre-training has a significant advantage in classification performance;compared with gray-scale image training,the performance of VGG-GAP trained by images after pseudo-color preprocessing was improved by about 5%.3)To solve the problem that the two-dimensional image segmentation method lacks spatial context information on the three-dimensional medical image,a segmentation network named CLUNS based on the convolutional recurrent neural network was proposed.The experimental results show that the CLUNS network with spatial context learning ability has reached the level of manual segmentation in the three-dimensional semantic segmentation of the thyroid;comparing the segmentation results of the two-dimensional network,CLUNS effectively reduces false positives and false negatives,and the Dice score was improved from 0.92 to 0.94.This article opens up a new research direction for the computer-aided diagnosis of thyroid cancer,and has achieved preliminary results.Several related technologies,such as image segmentation and deep pre-trained models,have great potential for development and application.With the accumulation of images,the method in this paper can achieve very good results,at the same time it is more flexible and convenient,and has practical value for reducing the misdiagnosis rate of thyroid cancer.