Classification Of Gastric Cancer Lesions In CT Images Based On Deep Learning

The early diagnosis of gastric cancer is closely related to the prognosis of patients.If the pathological status information of gastric cancer can be accurately evaluated,clinicians can carry out personalized treatment according to the actual situation of patients,which can more effectively evaluate the prognosis of patients.The traditional early diagnosis of gastric cancer mainly depends on pathological biopsy,which is not only limited by the clinical experience of gastroscope operators,but also causes invasive damage to patients.In recent years,CT has become an important auxiliary cancer diagnosis technology.Doctors can make a preliminary judgment on cancer with the help of CT images.Studies have proved that the texture analysis of CT images has important reference value for the diagnosis of the degree of differentiation of gastric cancer.The pathological analysis of gastric cancer on CT images using deep learning method provides a non-invasive method for cancer diagnosis.The actual clinical classification tasks of medical images often require the model to have high feature expression ability.At this stage,the deep learning model has limited feature expression ability for CT images of gastric cancer,and the classification effect is not ideal.Meanwhile,it also depends on a large number of data sets and accurate annotation information.Therefore,this paper studies the classification of differentiation degree of gastric lesions based on CT images.Aiming at the problems of limited feature expression ability and few gastric CT dataset,this paper proposes a gastric cancer lesion classification model Aux-IDPNet(Auxiliary training and Improved-DPN Network)based on Res Next and a multi classification model SSP-PSSANet(Self-Supervised Pre-training with Pyramid Squeeze and Spatial Attention Network)based on the idea of self supervised learning and attention mechanism.Using the above model,a gastric cancer lesion classification system based on CT image is developed for clinicians’ auxiliary diagnosis.The work of this paper mainly includes the following four aspects:(1)This paper proposes a preprocessing method based on multimodal data.In order to better characterize gastric tumors and improve the classification effect,this paper effectively preprocesses CT image data,extracts label information from the clinicopathological state information provided by doctors,explains in detail the extraction of weak supervised annotation,and finally obtains the dataset for training.(2)For the dataset with weak supervised annotation,the Aux-IDPNet model is proposed in this paper.The model introduces the attention map generator and improves the DPN structure,so as to improve the repeated mining ability of the model for CT image features,and then learn more new features; The auxiliary training branch is designed and added,and the weak supervised annotation of doctors is introduced as external learning information to improve the accuracy of the model.(3)Aiming at the problem of few public dataset of gastric medical images and class imbalance,this paper proposes SSP-PSSANet model to realize multi classification of the differentiation degree of gastric cancer lesions.The model integrates the idea of self supervised learning.Firstly,the label information is abandoned for training,and then the pre training model is used for supervised training to strengthen the classification effect,which effectively makes up for the problem of few CT data sets and unbalanced categories at the present stage; At the same time,a new PSSA attention module is proposed,so that the model can more effectively extract the spatial information of different scales,form the channel dependence of long distance,more accurately combine the context features of adjacent scales,make the model pay more attention to the tumor part of gastric cancer,mine more hidden information contained around,and then improve the classification effect of the model.(4)Based on the above model research,this paper designs and implements a medical assistance system for gastric cancer lesion classification,which realizes the reading of medical CT image data,slice display,prediction of classification results,and generates a medical result report on gastric cancer under the confirmation of doctors,which provides effective medical assistance for doctors.In this paper,115 dataset with differentiation labels and 208 dataset with differentiation labels were tested,and AUC of 0.8135 was obtained in two classifications and F1-scores of 0.7372 in multiple classifications,which are better than the control method,and can better distinguish ”differentiated” from ”undifferentiated” and tumor differentiation.The effectiveness of the proposed algorithm is proved.