The Segmentation And Recognition Of Oral Leukoplakia And Lichen Planus Based On Multi-task Learning

Oral leukoplakia refers to white or grayish-white plaque lesions on the oral mucosa.Oral lichen planus is a common chronic oral mucosal disease,both of which belong to the category of non-infectious oral mucosal diseases.Among them,oral leukoplakia belongs to the category of precancerous lesions or potential malignant diseases.At present,the diagnosis of oral mucosal diseases mainly depends on the clinical experience of doctors,and the diagnosis results are susceptible to subjective factors,and the diagnosis efficiency is low.For this reason,this paper mainly studies the auxiliary diagnosis method of automatic recognition of oral mucosal diseases based on computer vision technology.Before,the auxiliary diagnosis method usually uses the training model of single-task learning neural network to mine the features in the data for true and false discrimination.Aiming at the problem of small amount of medical data,this paper studied the recognition method of oral mucosal diseases based on data augmentation strategy,and further proposed the classification and recognition model of oral diseases based on multi-task learning.The main research contents are:1.In order to solve the problem of intra-class unbalance of oral leukoplakia and lichen planus,this paper proposes a sparse deep convolution generative adversarial network method is proposed to enhance unbalanced data.The generative model generates a new oral image and the discriminative model can judge whether the image is real oral mucosal lesions data or not.Through the back propagation,the generative model and the discriminative model confront each other.The generative model can generate more lifelike oral samples as much as possible,and the discriminative model can judge the authenticity of the samples as much as possible.Both of them are continuously optimized and improved to get the optimal model.The paper adds the sparse limitation optimizes the generative model to improve the dropout method,calculate the value of node activation,set the neuron node to 0 from small to large,keep the larger activation value node can make the most of the effective information of the image.The experimental results show that this method can effectively solve the problem of poor classification results caused by the imbalance of oral data,and the classification accuracy is improved by 6.30%.2.To solve the problem that the images of oral leukoplakia and oral lichen planus have large intra-class differences but small inter-class differences,this paper proposes a segmentation and recognition method based on multi-task learning.Firstly,the image is preprocessed and resized,and pretraining model is used as the starting point of the multi-task learning model to migrate to the target task.Then,the convolutional neural network is used to extract the universal feature map information on the basis of the shared network,specific layer is designed for each task to extract the feature map more discriminable.Finally,the detection and the segmentation network give the results by sharing the characteristic feature map of the network and their own feature map.The classification subnetwork combines the characteristic map of the segmentation network to segment and identify the oral leukoplakia and lichen planus together.Experiments show that this method can effectively improve the classification accuracy,and can segment the lesion area sensitively and accurately.Classification accuracy and segmentation results were improved by 4.25% and 12.98%.