Intelligent Segmentation And Staging Diagnosis Of Esophageal Cancer Based On Deep Learning

Objective:Esophageal cancer(EC)originates from the mucosal layer of the esophagus and has the seventh highest incidence rate in the world.It is difficult to be diagnosed at an early stage,with rapid development and poor prognosis.The treatment of esophageal cancer varies according to the stage of the disease,mainly based on surgery,and the surgical method is mainly based on TNM stage,with direct surgery at T1-T2 stage and surgery after neoadjuvant radiotherapy at T3-T4 stage,so preoperative T-stage diagnosis of esophageal cancer is the key to choose the treatment for all patients.Clinically,the preoperative diagnosis and TNM staging of EC mainly rely on computed tomography(CT).In order to solve the difficulties of early diagnosis of EC,this study proposes to construct a 3D model of esophageal cancer based on chest-enhanced CT and 3D reconstruction technology,create a measurement database of relevant anatomical structures of esophageal cancer and its adjacent structures,and use statistical analysis of anatomical morphological data of relevant esophageal cancer and its adjacent structures to explore the anatomical morphological differences of each T-stage of esophageal cancer and the correlation with T-stage,which is expected to assist clinicians to make T-stage diagnosis based on morphological parameters such as volume,length,and width of esophageal cancer.In recent years,with the rapid development of deep learning,many automatic segmentation models based on convolutional neural networks have been proposed and widely used for automatic recognition and segmentation of medical images.Since manual segmentation of esophageal cancer is an arduous and time-consuming task and prone to segmentation errors,automatic segmentation models for esophageal cancer based on CT images are urgently needed to be developed.In order to solve the current problem of insufficient artificial intelligence(AI)-assisted esophageal cancer(EC)enhanced CT image segmentation and staging diagnosis research,we plan to manually create EC segmentation datasets,conduct anatomical studies of T-staging of each esophageal cancer,and later construct deep learning-based intelligent segmentation model and intelligent T-stage diagnostic model to assist clinicians in diagnosis.The first part is the creation of deep learning annotated dataset of esophageal cancer and anatomical study of esophageal cancer T-stage,the second part is the creation of Uper Net-Swin based intelligent segmentation model of esophageal cancer and its adjacent structures,and the third part is the construction of intelligent diagnostic model of esophageal cancer T-stage based on enhanced CT images and Swin Transformer network.Methods:In the creation of a deep learning annotation dataset for esophageal cancer and anatomical study of T-staging of esophageal cancer,data from 155 thin-layer enhanced CT images of esophageal cancer were collected from two different hospitals,the First Affiliated Hospital of Army Military Medical University(Southwest Hospital)and Shanxi Cancer Hospital,from patients with postoperative histopathologically confirmed esophageal cancer.The segmentation dataset of esophageal cancer and its surrounding adjacent structures was created by manually segmenting the relevant structures including the tumor and its surrounding pericardium,thoracic aorta,lymph nodes,bronchus and lung using Amira 2019 software to create labels for a total of seven structures including the tumor.After that,measurements were performed on the created 3D model of esophageal cancer.The measurements included: surface area of esophageal cancer,volume of esophageal cancer,average long diameter of esophageal cancer,average short diameter of esophageal cancer,length of longitudinal axis of esophageal cancer,surface area and volume of normal esophagus,length of longitudinal axis of normal esophagus,and long and short meridians.Finally,the anatomical study of each T-stage of esophageal cancer was performed using Microsoft Excel and SPSS 25.0 software.In the study of creating an intelligent segmentation model of esophageal cancer and its adjoining structures based on Uper Net-Swin,a total of 182 cases of thin-layer enhanced CT image data of esophageal cancer were collected from two hospitals.In order to reduce the image errors caused by multicenter data,intensity normalization and image quantization were performed on the segmented region images by gray-scale range selection before training the model,so as to reduce the errors caused by image intensity variations.A deep learning network framework based on pytorch was used to create an intelligent segmentation model of esophageal cancer and its surrounding adjacent structures using Uper Net-Swin,Attention U-Net,Unet++,and Unet networks.The predictive performance of the model was evaluated using Dice similarity coefficient(DSC),sensitivity and positive predictive value(PPV)metrics.In the study of constructing an intelligent diagnosis model based on enhanced CT images and Swin Transformer network for esophageal cancer T-staging,150 cases of esophageal cancer thin-layer enhanced CT image data were collected from two hospitals.After image segmentation and preprocessing,they were input into three networks,Swin Transformer,Res Net50 and VIT,for EC T-stage intelligent diagnosis model construction.The model intelligent diagnostic results were evaluated using Precision,Recall,F1-score,Specificity,Negative Predictive Value(NPV),confusion matrix and Receiver operating characteristic(ROC)curve.Results:In anatomical studies of esophageal cancer,EC volume,long diameter,short diameter and EC surface area varied among T stages of esophageal cancer(AUC values for T2/T1,T3/T2 and T4/T3 were greater than 0.6),with higher values for long diameter and short diameter(AUC=0.761,0.675,0.709 and 0.775,0.744,0.748),and The risk level of different T stages was evaluated by calculating the critical values.the critical values of EC surface area,EC volume,long diameter and short diameter were 4008.50 mm2,11712.00mm3,24.25 mm and 14.69 mm for T2/T1,and the values of T3/T2 were 5533.50 mm2,19809.00 mm3,26.28 mm and 16.91 mm,and the values of T4/T3 were 10894.50 mm2,44103.50 mm3,37.05 mm,and 25.46 mm.In the intelligent segmentation study,the Uper Net-Swin model had the highest ECDSC value of 0.782,which was better than the Attention U-Net(DSC=0.6479),Unet++(DSC=0.6761)and Unet(DSC=0.6666)models,with the best segmentation effect,and the predicted values of EC volume and the volume values of manual segmentation The difference was not significant and there was no statistical difference between them(P >0.05).In the smart T staging diagnostic study,the Swim Transformer model had an AUC value of 0.861,much higher than the Res Net50 and VIT models of 0.611 and 0.542,with an accuracy of 1,0.67,0.83 and 1 for T1-4 staging.Conclusion:1.This study of 3D reconstruction of esophageal cancer based on enhanced CT images helps to accurately identify the location of esophageal cancer and to clearly and comprehensively observe the location,3D morphology and relationship of the tumor with adjacent structures.The tumor volume,long diameter and short diameter have a significant predictive effect on the diagnosis of T-stage of esophageal cancer,which helps thoracic surgeons and radiologists to make treatment and prognosis decisions.2.Using Uper Net Swin network,we successfully established an intelligent segmentation model for esophageal cancer,which can accurately segment esophageal cancer and its surrounding neighboring structures,outperforming Attention U-Net,Unet++,Unet and VUMix-Net.segmentation efficiency is no different from manual segmentation,and by accurately quantifying esophageal cancer volume,it facilitates accurate T-stage of esophageal cancer diagnosis and precisely guiding thoracic surgeons in surgical selection.3.Using Swin Transformer network,an intelligent T-stage diagnostic model for esophageal cancer was successfully constructed,which can diagnose the T-stage of esophageal cancer more accurately and enable clinicians to make early clinical interventions for patients.