Research On Image Detection Method Of Laryngeal Lesions Based On R-FCN

Clinically,a doctor usually diagnoses a laryngeal disease based on the image of a laryngeal lesion.In the process of looking at the image to find the lesion area,it usually consumes a lot of energy of the doctor.And may lead doctor to feel fatigued,resulting in low efficienty or misdiagnosis.Computer-assisted diagnosis of laryngeal lesions can avoid these problems.At present,computer-assisted laryngeal lesion image-assisted diagnosis systems mainly classify laryngeal lesion images.But this method is not good for doctors to check prediction results.In order to solve this shortcoming,this paper has designed an auxiliary diagnosis method that can give the lesion area and lesion category.And a lesion dataset is collected and created for the experiment.Aiming at the design requirements that the system needs to give the location and category of the lesion,a Region-based Full Convolutional Network(R-FCN)method for the detection of laryngeal lesions was proposed.In order to verify the recognition capability of the R-FCN model to recognize multiple organs,incomplete organs,and the organs whose shape changes due to lesions under complex backgrounds,we collect and create a laryngeal organ test set for experimental verification.The experimental results show that the average accuracy of R-FCN for epiglottis and vocal cords in organ test set is 93% and 91%,respectively.At the same time,the mean average precision(m AP)of the two is 92%.A 7-labeled laryngeal lesion data set was used to verify the effect of the difference in the number of object categories on the training effect of the R-FCN model.The experimental results show that with complex backgrounds and multiple targets in the image,the m AP detected by the R-FCN on the laryngeal lesion test set is 64.55%.Comparing the 92% m AP of the organ test set,it is shown that with the same number of images of each category,the increase in the number of categories of objects to be detected will reduces the model's detection accuracy.In addition,in the R-FCN test results for the category of vocal cord polyps in the laryngeal lesion test set,the size of most of the detection boxes is less than half the size of their corresponding labeled boxes..Aiming at the problem that the size of the vocal cord polyp detection box of the R-FCN model is significantly different from its labeled box.An improved method combining K-means algorithm with R-FCN is proposed.By analyzing the R-FCN detection box generation process and the commonly used clustering method,it is concluded that the size of the R-FCN detection box is related to the size of the anchor box.Based on this,we use the labeled box of the K-means clustering trainingset to obtain 9 rectangular box sizes,instead of the default size of the anchor box in the R-FCN model.The new R-FCN model was trained and tested with the laryngeal lesion data set.The experimental results show that the size of the vocal polyp detection box obtained by the improved method more closely matches the size of the labeled box in the dataset used in this paper.The m AP of the detection result on the laryngeal lesion test set increased from 64.55% to 69.01%.