| Rectal cancer is one of the most common digestive tract malignancies.Although the diagnosis and treatment of have been improved with the development of imaging and surgical procedures,the postoperative mortality rate remains high due to the high rate of recurrence and metastasis.Lymph node is one of the most major routes of rectal cancer metastasis and is also a key factor affecting postoperative survival.Mesorectal lymph nodes are the first affected and most frequently involved regional lymph node in rectal cancer.According to ESMO-2017（European society for Medical Onology）guidelines,as long as N+,regardless of T stage,the clinical stage is defined as Ⅲ or above,in which neoadjuvant chemoradiotherapy is required.Meanwhile positive circumferential resection margin（CRM）is an important independent risk factor affecting the recurrence after TME（Total mesorectal excision）surgery.The lymph nodes in close proximity to the mesorectal fascia may result in positivity of the CRM.In addition,N staging also determines rectal cancer risk indexing,which in turn affects the choice of treatment regimen.In summary,clear preoperative lymph node metastasis status is of great interest for the making of treatment protocols and reducing postoperative recurrence and metastasis.The clinical diagnostic methods for rectal tumors include MRI,CT,endorectal ultrasound,and pet-ct.MR examination,especially high-definition MR,is commonly used in the clinic and is also recommended by guidelines in various countries,allowing better preoperative assessment of local invasion of the tumor as well as metastasis of the mesorectal lymph nodes.The commonly used diagnostic criteria includes lymph node size,morphology,borders,signal within the node,etc.Malignant lymph nodes are generally larger than benign ones.However,studies have shown that size was not a very accurate indicator for the determination of benign and malignant lymph nodes.At the same time,internal signal and margin characteristics of lymph nodes can also be used to evaluate lymph node involvement.The malignant nodes signal tends to be uneven and to have irregular shape,lobulated,spiculated and indistinct margin.While the benign lymph nodes tend to have uniform internal signal,sharp and smooth edges and regular shape.However,although benign and malignant lymph nodes of rectal cancer have some diagnostic criteria according to the changes of morphology,border and internal signal,some studies have shown that the diagnostic efficacy of MRI in discriminating benign and malignant lymph nodes was unsatisfactory,with sensitivities and specificities ranging from 58%to 77%.At the same time,because of the relative subjective diagnostic criteria,it is difficult for physicians to make accurate and stable diagnosis.All in all,the clinic needs a subjective,stable,quantitative and simple method to diagnose malignant lymph nodes within mesorectum of rectal cancer.Radiomics refers to the acquisition of huge data by extraction of a large amount of information from medical imaging（CT,MRI,PET,etc.）.The main procedure of radiomics includes tumor segmentation,feature extraction and model building.By deep mining,prediction and analysis of the massive amount of imaging data,radiomics can assist physicians in making the most accurate diagnosis.It can overcome the limitations of human eyes in obtaining information and would not be affected by experience,time and even display equipment.Image segmentation is one of the most important process of the radiomics,and different segmentation methods have a great impact on the results.Deep learning is a branch of artificial intelligence.The deep learning system using multilayer convolutional neural network（CNN）can automatically extract,analyze quantitative image features and create prediction models.Compared with imaging radiomics,deep learning does not require accurate ROI（Region Of Interest）segmentation.It only needs to set the location and range of the lesion,and the computer can extract and analyze it by itself.Doctors do not need to participate in this process,which reduces the impact of subjective factors on the results.However,the data needed in deep learning is higher than radiomics to keep the training more stable.Broadly speaking,radiomics and deep learning both belong to the category of machine learning.Therefore,this study basing on high-definition T2 sequence images 1)used image radiomics combing different ROI segmentation methods to establish lymph node diagnosis model to discriminate metastatic and non-metastatic mesorectal lymph nodes of rectal cancer;2)attempted to use deep learning to predict the status of mesorectal lymph nodes in rectal cancer;3)evaluated the diagnostic value of machine learning,including radiomics and deep learning,and compared with manual diagnosis.The purpose of this study was to find a method that can objectively and quantitatively discriminate the benign and malignant mesorectal lymph node.Materials and methodsThe study was reviewed by the Shandong Qianfoshan hospital ethics committee.All examinations were agreed by the patients and the informed consent forms were signed.Patients with rectal lesions who came to Shandong Qianfoshan hospital from June 2016 to April 2021 were collected.The selected patients underwent MR examination before receiving any treatment and finally underwent TME surgery.MR examination used GE HD750 3.0T（GE Medical Systems,Boston,USA）magnetic resonance scanner and cardiac coil.The scanning sequence included high-definition T2WI with FOV was 18cm,thickness was 3mm and spacing was 0.In order to accurately make node-to-node correspondence between T2WI and pathological specimens,this study first partitioned and located the lymph nodes on the high-definition T2WI according to their relative position from the center of the tumor.After operation,the specimens were fully expanded and lymph nodes were searched according to the position of lymph nodes relative to the tumor center on T2WI.Upload the HD T2WI original DICOM images of lymph nodes to Huiyi Huiying radcloud workstation（Huiying Medical Technology Co.,Ltd,Beijing,China）.Four ROI segmentation methods were used to label lymph nodes,including contour,contour expansion,edge only and inner circle.In this study,1409 features were extracted,including intensity features,morphological features,texture features and high-order features.Firstly,the intra group and inter group consistency of features were judged,and the unstable features（ICC less than 0.75）were excluded.Variance thresholding、Select k Best and LASSO were used to reduce the dimension of the remaining features successively in order to increase the relevance of data.Finally,the features remained were included into the model fitting.The samples were randomly divided into training set and test set according to the ratio of 7:3.In the training set,the features were fitted to the logistics model,and 5-fold cross validation was used to increase data usage and improve the accuracy.The test set was used to verify the effect of the model.The software（platform）used was Python3.6（https://www.python.org/）.The selected images that can display the largest level of each lymph node were included in the deep learning study.The data sets were randomly divided into training set（70.0%）and testing set（30.0%）.A small CNN network with only 5 convolution layers and 2 full connection layers was designed.In addition,two data enhancement methods,random flip and adding random noise,were used to reduce the over fitting risk and improve the robustness of the model.According to the lymph node location,the VOI rectangular box can be obtained,and then the height value and width value of 1/4 can be expanded along the two directions of height and width on the basis of the rectangular box.Normalized the pixel value of each image,unified the image size,and then randomly flipped or added noise to the images.The optimizer used in this experiment was SGD（random gradient descent）.In order to reduce the over fitting risk,L2 regularization was also used.The software（platform）used was Caffe（Convolutional Architecture for Fast Feature Embedding）.At the same time,three doctors with different working experience independently read and diagnosed all lymph nodes without knowing the final pathological results.The short diameter of each lymph node was recorded at the same time.Statistical analysisSPSS（V20.0;IBM Corp.,Armonk,NY）and Medcalc 15.6.1（MedCalc Software Ltd.）software were used.The measurement data conforming to normal distribution were expressed as mean ± standard deviation（SX）.（1）Intra group correlation coefficient（ICC）was used to evaluate the consistency of ROI segmentation in imaging radiomics.ICC<0.4:poor consistency;0.40.75:the prompt has very good consistency.（2）The difference between the short diameters of benign and malignant lymph nodes was analyzed by t-test according to the results of variance,and Youden index was used to determine the optimal threshold.The evaluation indexes included ROC curve,AUC,accuracy,sensitivity and specificity.P<0.05 was the criterion of statistical difference.ResultsA total of 166 patients were enrolled.Male patients were more than female patients.T3 was the most in T stage.A total of 604 lymph nodes were matched between HD T2WI and specimens,including 306 malignant lymph nodes and 298 benign lymph nodes.The short diameter of lymph nodes were most below 6mm,accounting for 79.14%.The short diameter of positive lymph nodes was larger than that of negative lymph nodes,and the difference was statistically significant（P<0.05）.However,when the short diameter was used as the diagnostic standard,the AUC was 0.552.When the threshold was 5.93,the sensitivity was 0.307 and the specificity was 0.883.The result of radiomics.All 604 lymph nodes were divided into training set and test set according to the ratio of 7:3.The training set included 422 lymph nodes,including 215 malignant lymph nodes and 207 benign lymph nodes.The test set included 182 lymph nodes,including 93 malignant lymph nodes and 89 benign lymph nodes.The AUC,accuracy and sensitivity of the test set obtained by segmentation method 1 were the highest（0.820,0.725,0.756）.The specificity of the test set obtained by segmentation method 2 was the highest（0.772）.The results of deep learning.The data set was randomly divided into training（70.0%）and testing（30.0%）.Training set:215 positive+205 negative nodes,test set:91 positive+93 negative nodes.The results showed that the AUC of the test set was 0.81,the accuracy was 0.725,the sensitivity was 0.698 and the specificity was 0.752.The results of subjective diagnosis.The AUC,accuracy and sensitivity of the three doctors were not high（auc0.604,0.634,0.671,acc0.601,0.632,0.667,se0.366,0.552 and 0.392）,but the specificity was high（0.842,0.715 and 0.950）.The diagnostic efficacy of senior doctors was higher than that of low and middle-aged doctors,and the difference was statistically significant（P<0.05）.The diagnostic efficacy of middleaged doctors was higher than that of Low-aged doctors,but the difference was not statistically significant（P>0.05）.Conclusion1.The diagnostic efficiency was low when the short diameter of lymph node was used as the diagnostic standard,and it can not be used alone as the index for the diagnosis of mesorectal lymph node metastasis of rectal cancer.2.The method of ROI segmentation along the contour of lymph nodes finally obtained the best diagnostic efficiency,which means this method contains the most diagnostic information from the nodes.3.The diagnostic efficacy of deep learning and imaging radiomics was similar,and both were better than doctors’ subjective diagnosis.4.The results of subjective diagnosis showed that the working experience affected the diagnostic accuracy of benign and malignant lymph nodes.SignificanceImaging radiomics and deep learning based on HD T2WI can be used to distinguish and diagnose metastatic lymph nodes and non-metastatic lymph nodes in the rectum of rectal cancer,which is better than the subjective diagnosis of doctors.Machine learning can assist lymph nodes in the diagnosis of lymph node involvement in the mesorectum of rectal cancer,so as to provide more accurate condition of N-stage and circumferential margin,and provide useful information for the customization of personalized treatment and improve the prognosis of patients.Innovation1.Analyzed the application value of imaging radiomics and deep learning in the diagnosis of mesorectal lymph node metastasis in rectal cancer.At the present,there are few studies in this field.The node-by-node correspondence between image and pathology guaranteed the credibility of the results;2.Compared the effects of different ROI segmentation on the results of radiomics,which provides a theoretical basis for selecting a reasonable image segmentation method in the future research;3.Compared the diagnostic efficiency between machine learning and subjective diagnosis and proved that machine learning has a good potential application prospect in clinic.Deficiency1.The number of cases was small,and was a single center study.More number and multi center studies are needed for popularization and verification;2.Although the accuracy of lymph node localization can be improved by postoperative follow-up specimen processing,the structural changes and volume contraction of postoperative specimens would lead to the inaccuracy of localization,which is also a problem faced by all studies about the diagnosis of lymph nodes;3.No lymph nodes less than 3mm were included,and no lymph node micrometastasis was involved.At present,our diagnostic test can not reliably detect these lymph nodes. |
PDF Full Text Request