Diagnostic ability of artificial intelligence using deep learning analysis of cyst fluid in differentiating malignant from benign pancreatic cystic lesions
概要
Introduction
There are different types of pancreatic cystic lesions, such as intraductal papillary mucinous neoplasm (IPMN), mucinous cystic neoplasm (MCN), serous cystic neoplasm (SCN), and pancreatic pseudocyst (PPC). In IPMN, high-grade dysplasia or invasive carcinoma is often observed in 43% to 62% of patients who undergo surgical resection(Tanaka et al., 2012). IPMN with high-grade dysplasia or invasive carcinoma should be resected surgically, while IPMN with low- or intermediate-grade dysplasia should be monitored(Sahora et al., 2013). In MCN, invasive carcinoma is recognized in 4% to 16% of surgical resections(Baker et al., 2012; Jang et al., 2015; Yamao et al., 2011), and current guidelines recommend surgical resection(Tanaka et al., 2012). On the contrary, SCN and PPC are reported to have a slight risk of malignancy(Jais et al., 2016). Therefore, it is important to differentiate malignant from benign pancreatic cystic lesions to determine the appropriate treatment strategy.
However, it is difficult to differentiate malignant from benign pancreatic cystic lesions including SCN, PPC, epidermoid cysts (EDC), and lymphoepithelial cysts (LEC) based on clinical presentation and imaging modalities, using computed tomography (CT) and endoscopic ultrasound (EUS). Pancreatic cyst fluid analysis of several tumour markers, including carcinoembryonic antigen (CEA), carbohydrate antigen (CA) 72-4, CA125, CA19-9, and CA15-3, has been evaluated to determine whether the lesion is mucinous(Nagashio et al., 2014). The CEA of the pancreatic cyst fluid has been the most helpful tumour marker in differentiating mucinous from non-mucinous pancreatic cystic lesions.
The diagnosis of pancreatic cystic lesions remains challenging. An artificial neural network that imitates the cranial nervous system is a type of machine learning system, that is, artificial intelligence (AI). An artificial neural network consists of an input layer, a hidden layer, and an output layer. An artificial neural network with multiple hidden layers is called deep learning. This study aimed to investigate the diagnostic ability of carcinoembryonic antigen (CEA), cytology, and artificial intelligence (AI) by deep learning using cyst fluid in differentiating malignant from benign cystic lesions.
Methods
We retrospectively reviewed 85 patients who underwent pancreatic cyst fluid analysis of surgical specimens or endoscopic ultrasound-guided fine-needle aspiration specimens. For malignant and benign classifications of cystic lesions, high-grade dysplasia or invasive carcinoma (IPMN or MCN) was considered a malignancy, and low- or intermediate-grade dysplasia (IPMN or MCN), PPC, EDC, and LEC were considered benign lesions.
A neural network is a machine learning algorithm that provides mathematical result after entering various numerical values or information. In this study, to construct an AI-based diagnostic algorithm, deep learning, a multi-hidden layer of neural network. AI using deep learning was used to construct a diagnostic algorithm. TensorFlow version 1.5 (Google LCC, Mountain View, USA) was used for the deep learning analysis. During the training process of deep learning, labelled information such as clinical information was entered into the deep learning algorithm. CEA, carbohydrate antigen 19-9, carbohydrate antigen 125, amylase in the cyst fluid, sex, cyst location, connection of the pancreatic duct and cyst, type of cyst, and cytology were keyed into the AI algorithm, and the malignant predictive value of the output was calculated. The primary endpoint was to investigate and compare the diagnostic ability of AI, AI using only CEA, cyst fluid analysis (CEA, CA19-9 CA125, and amylase) and cytology in differentiating malignant from benign pancreatic cystic lesions.
Results
A total of 138 patients underwent pancreatic cystic lesions analysis, of which 53 patients with missing data were excluded. Eighty-five patients were analysed in this study. The mean age was 58.2 ± 13.4 years, 35 were men, and 50 were women. The final diagnosis of 23 patients was malignant and that of 62 patients was benign.
Area under receiver-operating characteristics curves for the diagnostic ability of malignant cystic lesions were 0.719 (CEA), 0.739 (cytology), and 0.966 (AI). In the diagnostic ability of malignant cystic lesions, sensitivity, specificity, and accuracy of AI were 95.7%, 91.9%, and 92.9%, respectively. AI sensitivity was higher than that of CEA (60.9%, p=0.021) and cytology (47.8%, p=0.001). AI accuracy was also higher than CEA (71.8%, p<0.001) and cytology (85.9%, p=0.210).
The univariate and multivariate analyses of malignant cystic lesion including AI, pancreatic cyst fluid, and clinical data are performed. In the univariate analysis, CEA, connection of the pancreatic duct and cyst, and AI were statistically significant. Variables with p < 0.05 in the univariate analysis were selected for entry into the multivariate analysis. In the multivariate analysis, CEA (p = 0.036; odds ratio [OR], 31.5; 95% confidence interval [CI], 1.3-794.4) and AI (p < 0.001; OR, 177.2; 95% CI, 12.0-2612.2) were significant factors in the diagnosis of malignant cysts.
Discussion
Diagnosis of malignant pancreatic cystic lesions is necessary for determining the appropriate treatment strategy. The diagnostic ability of CEA using pancreatic cyst fluid for malignant cysts was poor, and a previous meta-analysis reported its diagnostic sensitivity of 63% and specificity of 63%(Ngamruengphong et al., 2013). Even in the present study, the diagnostic ability of CEA by cyst fluid analysis was poor. Several studies have noted elevated cyst fluid CEA in LEC (Raval et al., 2010) and EDC(Higaki et al., 1998). These studies reported CEA expression on the surface of squamous epithelial cells, indicating that squamous epithelial cells could produce CEA, resulting in the elevated cyst fluid CEA levels. In our study, two patients had EDC and two patients had LEC. These results increased the cut-off value of CEA in differentiating malignant from benign pancreatic cystic lesions. This is an important limitation of differentiating malignant from benign cystic lesions using cyst fluid CEA.
AI using deep learning is used for various fields such as skin cancer(Esteva et al., 2017), radiation oncology(Bibault et al., 2016), Helicobacter pylori infection(Shichijo et al., 2017), colorectal polyp(Byrne et al., 2017; Chen et al., 2018), and histopathology(Litjens et al., 2016) using image information. However, the application of AI using deep learning to analyse pancreatic cystic lesions is relatively new. We used a system of AI using deep learning to simulate the human brain nervous system. In this study, we constructed AI using deep learning with TensorFlow. This framework accepts sets of data and corresponding labels as input clinical data and constructs a neural network for diagnosis. In this study, AI using deep learning analysed pancreatic cyst fluid and clinical data. By using this deep learning method, AI learns the characteristics of malignant cystic lesions by combining cyst fluid analysis and clinical data, and AI can possibly exclude the bias generated by human judgment. Although it is difficult for clinicians to diagnose malignant pancreatic cystic lesions by cyst fluid analysis and clinical data, AI using deep learning achieved adequate diagnostic ability in differentiating malignant from benign cystic lesions compared to cyst fluid analysis such as CEA and cytology. AI and CEA were also significant factor in the multivariate analysis of malignant cystic lesion. Specifically, although it is generally a problem that cytology diagnosis has low sensitivity, AI using deep learning achieved high sensitivity (95.7%). AI could be a powerful tool for exclusion diagnosis of malignant pancreatic cystic lesions. Although AI using only CEA was slightly lower in diagnostic ability than AI, AI using only CEA has good diagnostic ability. If CA 19-9, CA 125, and amylase cannot be used in the cystic fluid analysis, AI using only CEA may be useful for diagnosis in differentiating malignant from benign pancreatic cystic lesions. AI will improve the diagnostic ability of pancreatic cystic lesions if introduced as a supporting system. In the future, an AI using deep learning-based diagnostic system will change the method of diagnosis of differentiating malignant from benign pancreatic cystic lesions.
AI may improve the diagnostic ability of differentiating malignant from benign cystic lesions. Moreover, compared to pancreatic cyst fluid analysis such as CEA and cytology, AI is highly sensitive of differentiating malignant from benign cystic lesions and may be useful for exclusion of malignant pancreatic cystic lesions.