Segmentation Algorithm Of Pleomorphic Adenoma Pathological Images Based On Deep Learning

Image segmentation of pleomorphic adenoma cells is an important tool in the treatment of tumours where the key to their subsequent treatment is early and accurate diagnosis,localisation and identification.At this stage,to accurately segment the histocytic lesioned areas of pleomorphic adenomas for better and more effective treatment,doctors must manually mark the images of pleomorphic adenomas,which poses two problems: firstly,it is time-consuming and inefficient,and in addition,the segmentation of pleomorphic adenomas requires a detailed determination of their different lesioned cell types,which requires not only a lot of energy but also more expertise.It also requires additional expertise.Secondly,different doctors may have different judgements on the same pathological image depending on their previous experience.This approach is highly subjective and may be less accurate when analyzing the specifics of pleomorphic adenoma.Recently,deep learning segmentation algorithms have found wide applications in medical image processing.In order to achieve automatic segmentation of pleomorphic adenoma images,in this paper,deep learning techniques are used as the main research method,combined with the characteristics of pleomorphic adenoma cells,validated on the pleomorphic adenoma segmentation dataset,and improvements are made to the UNet-based image segmentation algorithm in order to improve the segmentation performance of the network on pleomorphic adenoma pathological images and to achieve lesion region diagnosis of digital pleomorphic adenoma cell tissue.The main research efforts of the dissertation are as follows.(1)There is no authoritative public dataset available for the segmentation of pleomorphic adenoma images,this paper constructs a new dataset for the segmentation of pleomorphic adenoma by performing image annotation and pre-processing work for different pleomorphic adenoma image features through the acquisition of pleomorphic adenoma pathology image slides and learning related pathology knowledge from patients at Jilin University Stomatology Hospital.(2)By analysing the characteristics of FCN,UNet and Deeplab models and based on the training results,the classical UNet model in medical segmentation was selected as the base segmentation model in this paper for the segmentation of pleomorphic adenoma regions,and the feasibility of its automatic segmentation was verified through experiments.To solve the problem of poor segmentation accuracy in basic UNet model,a Dense-UNet network model is proposed,and the dense connection module is used to optimize the traditional UNet network,reducing the number of parameters in the network and improving the feature reuse rate.Moreover,because of the use of dense connection structure,it is more convenient for the network to be trained,the loss of gradient is reduced,and the accuracy of the segmentation is increased.In the image segmentation experiments of pleomorphic adenoma,Dense-UNet performed better than the traditional UNet in all evaluation metrics,with m PA metrics reaching 70.88% and m Io U metrics reaching 52.49%,an improvement of 6.81% in m PA metrics and 4.62% in m Io U metrics compared to the original UNet model.(3)Category imbalance is a common problem in image segmentation tasks,and this problem is more prominent in medical imaging,where the same category imbalance problem exists in the polymorphic adenoma image segmentation task.To reduce the influence of the unbalancing problem on the accuracy of image segmentation,this paper proposes a loss function BCEDice Loss for the segmentation of pleomorphic adenoma images by combining the balanced cross-entropy loss function based on cross-entropy and the overlap measure loss function based on the cross-merge ratio.BCEDice Loss achieved better experimental results.(4)To address the problems of missing feature information and low segmentation accuracy of UNet network in medical image segmentation,this paper proposes a Ras-UNet network model to further optimize the segmentation of histopathological images of pleomorphic adenoma.First,the concept of residual learning in ResNet is combined with the merits of the UNet model in the integration of deep and shallow features.To make up for the shortcomings of the UNet networker in terms of the lack of depth and accuracy in the representation of features,the traditional convolutional architecture is replaced by the residual structure.Secondly,as the number of layers in a medical image segmentation network increases,the problem of mixing valid and invalid feature information arises.Therefore,in this paper,different classes of histiocytes were distinguished from pleomorphic adenoma using the hybrid attention mechanism CBAM,which combines positional and channel attention to maximize the detection of key features in the network,provide interference suppression from irrelevant information,improve model learning efficiency,and capture connections between distant pixels.Ultimately,after the above improvements,the experimental results show that the Ras-UNet model achieves 73.16% in m PA metrics and 55.07%in m Io U metrics;compared to the Dense-UNet model,the m PA metrics improve by3.22% and the m Io U metrics improve by 4.92%,compared to the original UNet model,the m PA metrics improve by 10.25% and the m Io U metrics improve by 8.7 %compared to the original UNet model.The m Io U metric improved by 8.7%,and the segmentation results have improved significantly.