The Application Of Deep Learning-Based Radiomics In Diagnosis And Histopathological Risk Assessment Of Thymomas

Chapter 1 PrefaceThymoma is the most common thymic epithelial tumor.However,because of its low incidence rate,imaging studies are relatively few.It still faces great challenges in accurate diagnosis and histopathologic risk assessment of thymomas.Computed tomography(CT)has a high sensitivity for the diagnosis of anterior mediastinal tumors.It is an important examination method for the recognition and diagnosis of thymomas.The CT manifestations of thymomas and other tumors or tumor like lesions in the anterior mediastinum are often overlapped and difficult to distinguish,and it is also difficult to accurately distinguish low-risk and high-risk thymomas.Therefore,more accurate diagnosis and histopathological risk assessment of thymomas based on CT imaging is needed.Radiomics is a flow-based computer quantitative imaging analysis method.In the practice of conventional radiomics(CR),it is often necessary to manually sketch and segmentation of lesions and difficult to design algorithms to fully extract image features,so there are inherent problems and defects.At present,there are still few studies on radiomics of thymomas,and manual or semi-automatic methods are still used to segment thymomas,and the methods of features extraction and analysis of radiomics are limited.Quantitative analysis of tumor features or texture analysis are often used to identify their risk categories or evaluate their clinical stages,so the degree of intelligence is low and it is difficult to obtain comprehensive and more valuable information of lesions.Deep learning-based radiomics(DLBR)is a method of deep learning used in the process of CR,which is used for automatic segmentation and feature extraction of lesions.Finally,it constructs the models for classification and prediction of the clinical problems need to solve,which better makes up the defects in the stage of CR feature extraction.Therefore,DLBR in the study of thymomas is expected to improve the intelligence of radiomics and the accuracy of prediction results without increasing the additional sample size,which has great advantages and clinical application value.Chapter 2 CT imaging diagnosis and features analysis of thymomaObjective:To evaluate and summary the characteristic manifestations on CT images of thymomas and other tumors or tumor-like lesions in the anterior mediastinum.According to the pathological results of the patients,the CT imaging features of thymomas and non-thymoma were compared,and the relationship between CT imaging features and histopathological risk categories was further evaluated in patients with thymoma,so as to provide imaging basis for clinical diagnosis,treatment,and prognosis assessment.Materials and methods:A retrospective analysis was made on the two centers of China-Japan Friendship Hospital and Eastern Theater General Hospital.Three hundred and fifty-four patients who underwent anterior mediastinal mass resection from January 2011 to October 2019 had definite pathological results.All patients underwent chest contrast-enhanced CT within 2 weeks before operation,the reconstructed 5 mm slice thickness image combined with thin-slice images(0.5mm-1.25 mm)was used for this evaluation.CT images were reviewed retrospectively by two radiologists with 10 years’ working experience.The description criteria of CT features refer to the standard reporting terminology specification of anterior mediastinal masses by International Thymic Malignancy Interest Group(ITMIG),an international interest group on thymic malignant tumors.When two doctors do not agree on the imaging features,an imaging expert with 30 years’ experience will review them again and put forward the final opinion.According to the histological classification method of WHO,thymoma was histologically classified by combining the morphology of tumor epithelial cells and the ratio of lymphocytes to epithelial cells,which were divided into five different subtypes:type A,AB,B1,B2 and B3.They were divided into two categories:low risk(A,AB and B1)and high risk thymoma(B2 and B3)according to the different invasive biological behavior and prognosis of each subtype.Based on the evaluation of CT imaging features of all patients,the CT imaging findings of thymoma and non-thymoma were summarized and analyzed.Based on the CT imaging features of patients,the patients with thymoma and non-thymoma were compared in center 1 and center 2 respectively,and the relationship between CT imaging features and histopathological risk categories was analyzed between patients with low risk and high risk thymomas.Results:The lesion of type A thymoma is usually small,round or oval,with smooth edge,clear boundary and uniform enhancement density,while the lesion is highly suggested as type B3 thymoma by its large,irregular,lobulated,uneven enhancement density,cystic degeneration,necrosis and calcification in the tumor,unclear boundary,mediastinal fat infiltration and large blood vessel invasion,mediastinal pleural involvement or pleural effusion.The CT features of type AB,B1 and B2 thymomas overlap to some extent,so it is difficult to distinguish them accurately.Low-risk thymoma often needs to be differentiated from thymic cyst,thymic hyperplasia and mature teratoma,while high-risk thymoma needs to be differentiated from lymphoma,thymic adenocarcinoma and malignant germ cell tumors.There were significant differences in the number of MG and the contour of lesions between patients with thymoma and without thymoma,and there were statistical differences in the contour of lesions,vascular abutting≥50%,and pleural effusion between patients with low and high risk thymoma.Conclusions:Contrast-enhanced CT can clearly reflect the invasion or involvement of the mass itself and its surrounding tissues,and the preliminary diagnosis and differential diagnosis of thymoma can be made according to the findings of CT.The meaningful parameters and indexes can be provided for the diagnosis and risk assessment of thymoma based on comparative analysis of standardized CT features.Chapter 3 Conventional radiomics of thymomasObjective:Based on the contra-enhanced CT images of the chest,the flow and methods of conventional radiomics were used to construct a predictive model for the diagnosis of thymoma(Taskl)and histopathological risk assessment(Task2)combined with the clinical information of patients and the features of CT images.Materials and methods:A retrospective analysis was made on the two centers of China-Japan Friendship Hospital and Eastern Theater General Hospital.Three hundred and fifty-four patients whose thin-slices CT images(0.5-1.25mm)of mediastinal window were used in radiomics research underwent anterior mediastinal mass resection from January 2011 to October 2019 had definite pathological results.Two hundred and thirty-eight patients in center 2 were used for training dataset and 116 patients in center 1 were used for validation dataset in task 1 and 122 patients in center 2 were used for training dataset and 73 patients in center 1 were used for validation dataset.Two radiologists with 10 years working experience sketched the lesions manually,then the extracted the imaging features of the lesions were used to calculate the label value.Intraclass correlation coefficient(ICC)was used to evaluate the consistency of inter-observer and intra-observer.For Taskl and Task2,the imaging features of the lesions were extracted by the same method,and the Z value of the extracted features was standardized.The feature screening was carried out and the radiomics signature was calculated based on the ICC of extracted features,double-sample t-test and Least absolute shrinkage and selection operator(LASSO)regression.In the training dataset,radiomics signature combined with meaningful clinical information and CT imaging features were used as the final feature parameters.Multivariate Logistic regression analysis was used to build a predictive model for differential diagnosis between thymomas and non-thymomas and to distinguish the risk categories of thymoma.The Hosmer-Lemeshow test was used to test the goodness-of-fit of the evaluation model and the coincidence degree of the observed values in the training and the validation dataset.Results:Two doctors sketched the lesions manually,the ICC values within the sketchers were 0.911(95%CI:0.822-0.957)and 0.897(95%CI:0.785-0.951),respectively and the ICC values between the sketchers were 0.852(95%CI:0.713-0.927).The number of extracted image features was 1332,and two features were retained to calculate the radiomics signature by LASSO regression in Taskl and Task2.The multivariate Logistic regression analysis was used to construct the prediction model,in which the variables of the Task1 prediction model included MG,contour and radiomics signature.The AUC of the prediction model in the training dataset was 0.7936(95%CI:0.7368-0.8503),the sensitivity was 0.7586,the specificity was 0.7213,and the accuracy was 0.7395.The model was tested by HL test and P is 0.9015.The AUC was 0.7025(95%CI:0.6041-0.8007)in the verification dataset,the sensitivity was 0.8082,the specificity was 0.4419,and the accuracy was 0.6724.The model was tested by HL test and P is 0.5261.The variables of the Task2 prediction model included contour,vascular abutting≥50%,pleural effusion and radiomics signature.The AUC of prediction model in the training dataset was 0.8094(95%CI:0.7298-0.8890),the sensitivity was 0.8873,the specificity was 0.5686,and the accuracy was 0.7541.The model was tested by HL test and P is 0.7124.The result of Task2 in the validation dataset showed that AUC was 0.7930(95%CI:0.6699-0.9161),the sensitivity was 0.9400,the specificity was 0.5217,and the accuracy was 0.7082.The model was tested by HL test and P is 0.8230.Conclusions:Conventional radiomics provides an objective and quantitative method for distinguishing thymoma and its histopathological risk categories,which is helpful to assist clinical diagnosis and treatment of thymoma.Chapter 4 Automatic Segmentation of anterior mediastinal lesions based on Deep LearningObjective:To segment the lesions automatically by deep learning method based on the chest contra-enhanced CT images and the manual sketch results of all anterior mediastinal lesions,and the results of segmentation were evaluated.Materials and methods:A retrospective analysis was made on the two centers of China-Japan Friendship Hospital and Eastern Theater General Hospital.Three hundred and fifty-four patients whose thin-slices CT images(0.5-1.25mm)of mediastinal window were used for automatic segmentation underwent anterior mediastinal mass resection from January 2011 to October 2019 had definite pathological results.Two hundred and thirty-eight patients in center 2 were used for training dataset and 116 patients in center 1 were used for validation dataset.Deep learning method was used to segment the anterior mediastinal lesions automatically based on the results of manual sketching of all lesions.The segmentation methods included the design of double lung mask files to remove the mediastinal area,the initial segmentation by V＿Net network and following the accurate segmentation by morphological snake algorithm.After the segmentation is finished,the segmentation results are evaluated by three indicators:dice coefficient,accuracy and recall.Results:The average dice coefficient,precision and recall in the training dataset were 0.942±0.066,0.915±0.083 and 0.907 10,091,respectively.The average dice coefficient,precision and recall in the validation dataset were 0.911±0.051,0.926±0.042 and 0.89±0.059,respectively.The dice coefficient curve after automatic segmentation of lesions showed that in the training dataset,the minimum value of dice coefficient was 0.623 and the maximum value was 0.999,and the dice values of 125 samples were more than 0.95(52.51%).In the validation dataset,the minimum value of dice coefficient was 0.624,the maximum value was 0.998,and the dice values of 78 samples were in the range of 0.85 to 0.95(67.24%).Conclusions:The effect of accurate segmentation of anterior mediastinal lesions by using V-Net network combined with morphological snakes algorithm is better,which can meet the needs of further deep learning quantitative feature extraction and analysis,which provides a methodological basis for the research of automatic segmentation of anterior mediastinal lesions.Chapter 5 Deep learning-based radiomics of thymomasObjective:Based on the contra-enhanced CT images of the chest,the flow and methods of Deep learning-based radiomics(DLBR)were used to construct a predictive model for the diagnosis of thymoma(Taskl)and histopathological risk assessment(Task2)combined with the clinical information of patients and the features of CT images.Materials and methods:A retrospective analysis was made on the two centers of China-Japan Friendship Hospital and Eastern Theater General Hospital.Three hundred and fifty-four patients whose thin-slices CT images(0.5-1.25mm)of mediastinal window were used in radiomics research underwent anterior mediastinal mass resection from January 2011 to October 2019 had definite pathological results.Two hundred and thirty-eight patients in center 2 were used for training dataset and 116 patients in center 1 were used for validation dataset in task 1 and 122 patients in center 2 were used for training dataset and 73 patients in center 1 were used for validation dataset.After automatic segmentation,the deep learning method was used to extract the features of lesions.The deep learning network was trained with ResNet-34 containing two residual blocks and the deep learning features are extracted from the first full connection layer.The features were screened by Kendall correlation coefficient,10-fold cross-validation method and LASSO regression.Finally,multivariate Logistic regression analysis is used to construct prediction models.Hosmer-Lemeshow test was used to test the goodness-of-fit of the prediction results in the training and the validation dataset.Delong test was used to compare the DLBR predictive results with the results of conventional radiomics.Results:In Taskl and Task2,the number of features is 4096 based on the number of neurons in the first full connection layer of ResNet,and three radiomics features were retained for calculating radiomics signature by LASSO regression in Taskl and Task2,respectively.The multivariate Logistic regression analysis was used to construct the prediction model.The variables of the Taskl prediction model were MG,contour and DLBR signature.The AUC of the prediction model was 0.8344(95%CI:0.7833-0.8856),the sensitivity was 0.8017,the specificity was 0.7309,and the accuracy was 0.7605 in training dataset.The HL test showed that P value is 0.8867.The AUC was 0.7415(95%CI:0.6496-0.8334),the sensitivity was 0.7808,the specificity was 0.5581,and the accuracy was 0.6983 in the validation dataset.The HL test showed that P value is 0.6547.The variables of Task2 prediction model included contour,vascular abutting>50%,pleural effusion and DLBR signature.In the training dataset,the AUC of the prediction model was 0.8255(95%CI:0.7454-0.9055),the sensitivity was 0.9014,the specificity was 0.6078,and the accuracy was 0.787.The HL test showed that P value was 0.9033.In the validation dataset,the AUC of the prediction model was 0.7752(95%CI:0.6512-0.8992),the sensitivity was 0.9000,the specificity was 0.6087,and the accuracy was 0.8028.The HL test showed that P value was 0.5434.The Delong test was used to compare the ROC prediction results between DLBR and CR in Taskl and Task2 respectively in the training dataset.In Taskl,the AUC of DLBR model was 0.8344(SE=0.0261),CR model was 0.7936(SE=0.0290),the difference was 0.0409(95%CI:0.0001-0.0816)and was statistically significant.In Task2,the AUC of DLBR model was 0.8255(SE=0.0415),CR model was 0.8094(SE=0.0406),the difference was 0.0146(95%CI:-0.0242-0.0534)and was not statistically significant.According to the decision curve analysis,DLBR model had more clinical benefits than CR model when the threshold probability is greater than 46%in Taskl;DLBR model had more clinical benefits than CR model when the threshold probability is greater than 60%in Task2.Conclusions:In the histopathological risk assessment of thymomas,DLBR can be used to replace the steps of segmentation and features extraction in the workflow of conventional radiomics without the additional increasing sample size.Compared with CR,DLBR can improve the intelligence of diagnosis and assessment process,and has better prediction effect as well as greater clinical benefits.