| Part Ⅰ1 A Large Sample Control Study of Esophageal Wall Thickness on Preoperative CT and Pathological T Stage of Esophageal CancerObjectiveTo explore the correlation between esophageal wall thickness(EWT)on preoperative CT and postoperative pathological T stage of esophageal cancer(EC),and determine the optimal EWT threshold for the T stage.Materials and methods1102 consecutive patients with histopathologically confirmed esophageal squamous cell carcinoma(ESCC)between July 2014 and April 2020 were retrospectively reviewed.All patients underwent preoperative contrast-enhanced CT and surgical treatment.Patients were divided into pT1,pT2,pT3 and pT4 groups according to the pathologic stage.The maximal EWT of the lesions on CT was measured.The EWT of each group was calculated and compared using independent samples t-test and one-way analysis of variance(ANOVA).90%of samples from each subgroup were randomly selected as the training set,while the remainder comprised the testing set.We employed the support vector machine(SVM),where linear kernels were leveraged to determine the optimal threshold to classify samples with different T stages..Results1.The mean EWTs of the pTl,pT2,pT3 and pT4 groups were 4.9±2.6 mm,8.1±2.3 mm,12.4±3.6 mm,and 18.6±4.4 mm,respectively.Differences in the EWT between the four groups or between adjacent groups(pTl and pT2,pT2 and pT3,pT3 and pT4)were significant(P<0.001),and esophageal wall became thicker with increasing pT stage(pT1<pT2<pT3<pT4).2.Through the construction and operation of the SVM model,the results showed that the optimal thresholds between samples from pTl/pT2,pT2/pT3 and pT3/pT4 lesions were 5.5 ± 0.3 mm,10.8±0.8 mm and 15.9± 0.5 mm,respectively.The accuracy of EWT for T-staging prediction was 60.29±2.33%.ConclusionsThe EWT on preoperative CT is closely related to pathological T stage,which can be used to predict the T stage of ESCC.Through the model examination with large sample size,the accuracy of T-staging prediction is 60%.2 The Value of Different CT Signs in T Stage of Resectable Esophageal CancerObjectiveTo explore the value of different CT signs in T-staging of resectable esophageal cancer(EC).Materials and methods1102 consecutive patients with histopathologically confirmed ESCC between July 2014 and April 2020 were retrospectively reviewed.All patients underwent preoperative contrast-enhanced CT and surgical treatment.Patients were divided into pTl,pT2,pT3 and pT4 groups according to the pathologic stage.The CT images were evaluated by two attending radiologists,including whether the lesions were visible,esophageal wall thickening,the circumferential ratio(CR),residual ring sign(RRS),circumferential sign(CS),outer contour and low enhancement sign of the outer wall.The CR refers to selecting the maximum level of lesion in the axial CT images,dividing the circumference of the esophageal wall into ten equal parts and calculating the proportion of lesions involving the esophageal wall,which ranges from 0 to 1.RRS refers to the CR is less than 1.CS refers to the CR is equal to 1.The smooth outer contour means that the outer outline of the esophageal wall is smooth within the scope of the lesion,otherwise the outer contour is not smooth.The enhancement of ESCC is often earlier than that of the normal esophageal wall on contrast-enhanced CT.The tumor tissue that has been enhanced in the early stage is compared with the outer wall of the esophagus that has not been fully enhanced.CT images show a difference in density between the inner and outer walls of the esophagus.If the outer wall of the esophagus with low enhancement is continuous and intact,it is defined as the low enhancement sign of the outer wall.Results1.The proportions of wall thickening in pTl,pT2,pT3 and pT4 ESCC patients were 16%(31/199),95%(128/135),100%(626/626)and 100%(142/142),respectively,and 4%(4/100)and 27%(27/99)in pTla and pT1b ESCC patients,respectively.2.In ESCC patients with wall thickened,the proportions of RRS in pTl,pT2,pT3 and pT4 ESCC were 100%(31/31),80%(102/128),24%(153/626)and 9%(13/142),respectively.The proportions of CS in pTl,pT2,pT3 and pT4 ESCC was 0%(0/31),20%(26/128),76%(473/626)and 91%(129/142),respectively.3.The occurrence probability of RRS in pT1a,pTb and pT2 ESCC were 4%(4/100),27%(27/99)and 76%(102/135),respectively.The CR of the pT1a,pTb and pT2 ESCC were 0.4±0.1,0.4±0.1 and 0.6±0.2,respectively.4.The proportion of smooth outer contours in pT3 and pT4 ESCC patients was 63%(394/626)and 20%(28/142),respectively.The esophageal contours of pTl-2 ESCC patients were smooth.The sensitivity and specificity of non-smooth outer contour in the diagnosis of fibrous membrane invasion of ESCC were 0.45 and 1,respectively.5.The proportion of low enhancement of the outer wall in pT1a,pT1b and pT2 ESCC was 50%(3/6),45%(22/49)and 27%(36/131),respectively,which was not found in patients with pT3-4 ESCC.The low enhancement of the outer wall was mainly seen in the arterial phase,accounting for 90%(55/61).Conclusions1.Whether the esophageal wall thickens or not is an important basis for distinguishing pT1 from pT2-4 ESCC.2.The RRS highly indicates that the lesion is in stage T1-2.Since T1-2 ESCC can show the RRS,T1-2 ESCC can be distinguished from the occurrence probability of the RRS and the CR.The probability of the RRS and the CR of the lesions in pTla and pT1b ESCC are less than that in pT2 ESCC.3.The CS indicates that the lesions are most likely to be in stage T3-4.4.The specificity of non-smooth outer contour in the diagnosis of fibrous membrane invasion of esophageal carcinoma is high,but the sensitivity is low,only 0.45.5.The low enhancement of the outer wall is of great value in the staging of pTl-2 ESCC,which mainly occurs in the arterial phase.CT images in the arterial phase should be under the spotlight for the diagnosis and staging of early EC.3 The Detection Rate and CT Findings of Esophageal Cancer in Different T StagesObjectiveTo investigate the detection rate and imaging findings of EC with different T stages on contrast-enhanced CT.Materials and methods1102 consecutive patients with histopathologically confirmed ESCC between July 2014 and April 2020 were retrospectively reviewed.All patients underwent preoperative contrast-enhanced CT and surgical treatment.Patients were divided into pTl,pT2,pT3 and pT4 groups according to the pathologic stage.To evaluate the detection rate of ESCC with different T stages on contrast-enhanced CT.At the same time,by analyzing the commonness of ESCC with the same T stage on CT images,so as to summarize the CT findings of T1-4a EC and clear the value of contrast-enhanced CT in T-staging of EC.Results1.The detection rates of pTl,pT2,pT3 and pT4 ESCC were 28%(55/199),97%(131/135),100%(626/26)and 100%(142/142),respectively.The detection rates of pTla and pT1b tumors were 6%(6/100)and 50%(49/99),respectively.It is worth noting that the general types of pT1 ESCC that can be detected are mainly mushroom type,followed by protuberant type,while the general types undetected are mainly superficial and concave type.2.The CT findings of each stage ESCC were as follows:(1)There was no thickening of esophageal wall in pT1 ESCC(84%);the characteristic manifestation of pT1a ESCC was no thickening of esophageal wall and the lesion was invisible(94%);pT1b ESCC was mainly invisible(51%),followed by the RRS(27%)and abnormal enhancement of mucosal surface(22%).(2)The main manifestations of pT2 ESCC were mild thickening of esophageal wall(95%),the RRS(76%)was typical feature and it also appeared mild CS(19%).(3)The main manifestations of pT3 ESCC were thickening of esophageal wall with varying degrees(100%),and the CS(76%)was typical feature.The outer contours of esophagus could be smooth(63%)or not(37%).(4)The main manifestations of pT4 ESCC were circumferential wall thickening,non-smooth outer contour and the invasion of adjacent pericardium,pleura,azygos vein,diaphragm and so on.Conclusions1.The detection rate of pT1 ESCC is low,and the gross classification of lesions is the main factor affecting whether pTl lesions can be detected by CT.2.The main points of differentiation of ESCC with different T stages:(1)Most of pTl and pT2 ESCC can be distinguished on CT images,the main basis for distinguishing between them is esophageal wall thickening or not,no thickening wall indicates pTl stage,thickening wall indicates T2 stage,a few pT1b ESCC with mushroom and protuberant type and pT2 ESCC are not easy to distinguish.(2)The differentiation of pT2 and pT3 ESCC is mainly based on the RRS and the CS,the RRS indicates T2 stage,the CS indicates T3 stage.(3)Both T3 and T4 mainly showed the circumferential sign,and the key to distinguish them was the relationship between the lesions and the structures.Part ⅡAutomatic Segmentation and T-Staging of Esophageal Cancer Using Deep Learning on CT ImagesObjectiveTo investigate the application value of the model based on deep learning algorithm in automatic segmentation and T-staging of EC on CT images.Materials and methodsCT datasets from 600 patients(mean age,63±8 years;477 male,123 female)with pathologically confirmed ESCC between July 2014 and February 2021 were retrospectively collected,and the ratio of the T1,T2,T3 and T4 ESCC patients was 1:1:1:1.The imaging datasets were randomly grouped using a ratio of 4:1.Dataset 1(480 cases)were used to construct the automatic segmentation model based on UNet and the T-staging model based on radiomics.Then,the segmentation model and the T-staging model were integrated to construct a two-stage automatic T-staging model.All primary tumors in dataset 1 were segmented manually.Dataset 2(120 cases)were used for testing,which were unprocessed original DICOM data and not exposed in the model training process.Performance of automatic segmentation model was evaluated by Dice coefficient(DSC)and mean Intersection-over-Union(mIoU).Performance of the AI T-staging tool was assessed by the receiver operating characteristic(ROC)curve and compared with the junior and senior radiologists.Results1.For the segmentation,experiments showed that the automatic segmentation model based on 3D-UNet outperforms that based on 2D-UNet(DSC,0.875±0.012 vs 0.846±0.007;mloU,0.907±0.014 vs 0.885±0.005).2.In staging task,pyradiomics software was used to extract a total of 827 features,and nine image omics features were screened out by the minimum absolute contraction and selection operator(LASSO)regression model,and the T-staging prediction model was further constructed using the selected features and SVM.The radiomics-based T-staging model in dataset 1 reached the areas under the curve(AUCs)of 0.94,0.81,0.92,and 0.97 for T1,T2,T3,and T4 ESCC,respectively.3.The end-to-end T-staging prediction of EC was realized through the integration of segmentation model and staging model,and the results showed that the AUCs of T1,T2,T3 and T4 ESCC were 0.94,0.76,0.74 and 0.88,respectively.The accuracy for T-staging by AI tool,junior radiologist and senior radiologist were 63.33%,53.33%,and 79.16%,respectively.ConclusionsThe model based on deep learning algorithm can carry out high-quality segmentation of EC.The end-to-end T-staging model has better diagnostic performance than junior radiologists,and AI has potential clinical application prospects in assisting the staging diagnosis of EC. |
PDF Full Text Request