Differentiation Of Gastrointestinal Submucosal Lesions Using Computer-aided Diagnosis Of Endoscopic Ultrasound(EUS)Images:a Diagnostic Test

Gastrointestinal submucosal lesions (SMLs) contain a variety of different pathological, for examples leiomyoma, gastrointestinal stromal tumors (GISTs). It is very difficult to diagnosis a SMLs using a conventional endoscopy because a normal mucosa is covered on the surface. Endoscopic ultrasonography (EUS) can clearly display the internal relationship and layer of origin of SMLs, echogenicity, heterogeneity and other information, and is a valuable imaging technique in the diagnosis of SMLs. However the diagnostic accuracy of SMLs using EUS is lower, especially for GISTs. GISTs are potential malignant tumors, and can spread to liver and peritoneal in early stage. Tumor remove is the only way to improve the prognosis of GISTs. Although the studies report that the sonographic characteristics of SMLs, such as location, size, layer of origin and echogenicity can help doctor diagnosis SMLs, but the sensitivity and specificity are controversial and the diagnosis criterions are also different. Recently, Endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA) has emerged as diagnosis method of SMLs. However, many factors have an impact on sample obtain rate of EUS-FNA and hasn’t enabled us to preoperatively diagnose the malignant GISTs. Therefore, it is a huge challenge for clinical endoscopists to find a noninvasive, objective and auxiliary method of EUS to diagnosis SMLs.Computer aided diagnosis (CAD) technology refers to the use of computer technology to extraction and analysis the texture features of EUS image, increasing the acquisition of image information, differential diagnosis lesion and pathological damage, further to achieve the purpose of differential disease. At present, many literatures reported that the CAD technology can be applied to diagnosis and classification of EUS images of various organs, including EUS images of SMLs. our team used to apply the CAD technology to extract the texture feature of pancreatic carcinoma and chronic pancreatitis EUS image, establishing a classification model and successfully separating EUS image of pancreatic carcinoma from EUS image of non-pancreatic cancer. Based on the above studies and CAD theory of our team, in this study, a CAD technology was used to differentiate the EUS image of SMLs.In this study, the EUS image of leiomyoma and four different risks GISTs (very low risk, low risk, medium risk and high risk) were involved in this study, and the viability of CAD technology to differentiate the EUS image of SMLs and separate benign and malignant EUS image of GISTs were proved. Because the current CAD techniques were used to classified two different samples, in the classification of GISTs risk, the EUS image of four different risks GISTs were divided into two categories:benign (very low risk and low risk) and malignant (intermediate and high risk) to establish two classification model in this study. Finally, the EUS image of four different risks GISTs was used to build a four classification model. This study contains follow two different parts:The first part:computer assisted image analysis technique in differential diagnosis of Ieiomyoma and GISTs.Objective:To investigate the diagnosis value of CAD in EUS image classification of GISTs and Ieiomyoma.Methods:180 EUS image of leiomyoma and 180 EUS image of GISTs were retrospectively collected in this study from 2000 January to 2013 December in database of Changhai Hospital endoscopy center. Select a typical lesion in EUS images, delineation the region of interests, interception the sub graph which is the largest rectangular of not more than the region of interests. A total of 157 parameters of 10 categories were extracted from the region of interest using computer-based techniques. ReliefF weighting method and sequential forward selection (SFS) were used to screening texture features for a better combination of features. Finally, a classification model based on support vector machine (SVM) was established, trained, and evaluated. Ten-fold cross-validation method was used to validate the classification performance of the model, the classification accuracy, sensitivity, specificity, positive predictive value and negative predictive value was statistical.Results:the best combination of texture features include 6 parameters of 3categories, in this stage, the classification accuracy rate reaches a maximum (75%). The best combination of texture features were used to establish classification model and 360 cases were classified, using ten-fold cross-validation validate the method, the final classification accuracy, sensitivity, specificity, PPV and NPV were 75.28%,77.26%,73.61%,74.83%,76.63%.Conclusion:computer aided diagnosis technology can be used to differentiate EUS image of GISTs and leiomyoma and provides a new research direction for the accurate diagnosis of gastrointestinal submucosal lesions.The second part:computer assisted image analysis technique in differential diagnosis of benign and malignant GISTs.Objective:To explore diagnosis value of CAD in EUS image classification of benign and malignant GISTs.Methods:the EUS image of four different risks GISTs which were diagnosed by pathological were retrospectively collected in this study from 2000 January to 2013 December in database of Changhai Hospital endoscopy center. the EUS image of four different risks GISTs were divided into two categories:benign (very low risk and low risk) and malignant (intermediate and high risk) to establish two classification model. Then, a classification model for four different risk GISTs was established. The texture feature extraction method and model establishment method were the same as described in Section one.Results:21 cases of very low risk GISTs,61 cases of low-risk GISTs,29 cases of intermediate risk GISTs and 13 cases of high risk GISTs were involved in this study from 2000 January to 2013 December. ReliefF weighting method and sequential forward selection (SFS) were used to screening texture features for a better combination of features. The best combination of texture features include 2parameters of 2 categories, in this stage, the classification accuracy rate reaches a maximum (75.9%).The best combination of texture features were used to establish classification model and 124 cases were classified, using ten-fold cross-validation validate the method, the final classification accuracy, sensitivity, specificity, PPV and NPV were 77.56%,94.03%,45.5%,77.27%,80%.Conclusion:computer aided diagnosis technology available to differential diagnosis of benign and malignant GISTs; further search for texture feature for EUS image classification of can improve the classification performance.Through the above study, this study draws the following conclusion:1. Computer aided diagnosis technique can be used for the diagnosis of SMLs, providing a new diagnostic method for clinical endoscopists to differential of leiomyoma and GISTs.2. The next step of the research is further search for texture feature for GISTs EUS image classification.