Application of AI to the molecular diagnosis of pancreatic cancer
In bioinformatics research, researchers often need to collect, screen, process, and summarize large amounts of data. As such, the question of whether machine learning can simplify the process and achieve good results has been a hot research topic. There are many specific molecules related to pancreatic cancer such as microRNA (miRNA) 10b, cell-free DNA, and ZIP4. Research on the molecular mechanism and diagnosis of pancreatic cancer has become a mature, standardized field, with a large number of relevant articles in recent years[11,12]. However, the need to collect and process data manually can consume a great deal of time and energy.
Machine learning helps researchers spend less time on data processing through one-time modeling. The steps for using machine learning typically include the following: Collecting the basic data, dividing data into an experimental group and a verification group, establishing a screening and processing model, inputting the experimental group data into the model, calculating the output results, and verifying the model’s feasibility using the verification group. The verification group can be used to test the specificity and sensitivity of the experimental group while the experimental group can make the model more intelligent. The steps are illustrated in Figure 1.
Figure 1 The common steps in machine learning.
A similar pattern was used for prognostic analysis of pancreatic cancer.
Using network representation learning and convolutional neural networks, the correlation between miRNA and pancreatic cancer disease can be analyzed, and the potential disease miRNA can be found. Machine learning has been used to process large exocrine RNA data and generate predictive templates that can identify cancer in individuals. An ANN can imitate the human neural meridian system. It is divided into three parts: Input layer, hidden layer, and output layer. “Deep learning” (Figure 2) refers to an ANN with multiple hidden layers. Using this technique, cyst tumor markers, amylase, cytology, and other information are inputted and then combined with two data; the output layer outputs whether the pancreatic cystic lesions are benign or malignant. Some researchers have also proposed an extensible supervised classifier technical framework that can diagnose pancreatic cancer provided the expression profile of a single cell can be input to reveal its identity.
Figure 2 Algorithm of deep learning.
Data were passed from layer to layer: from the input layer to the output layer.
Although machine learning can save researchers a lot of time on data processing, machine learning still has many limitations. The first concerns data collection and processing. Specific input projects at the beginning of modeling are needed for machine learning and neural network analysis. However, for researchers who have not carried out data analysis, it is unknown which raw data are necessary and unnecessary. Useless data simply increase the workload and can also become the specificity and sensitivity of the model. Meanwhile, editing the model also poses a significant problem. Although AI can save time, the threshold and workload in the establishment of AI programs are prohibitive for nonprofessionals who lack a foundation in math and programming.
The occurrence and development of pancreatic cancer is complex and changeable, and the patient’s condition has a large degree of variability. In this regard, AI can be applied to the molecular diagnosis of pancreatic cancer and can obtain objective data processing results. However, AI is not independent and mostly can only be used as an auxiliary tool. Yet, with continuous development and improvement, AI might eventually have a more universal application.
Application of AI in the imaging diagnosis of pancreatic cancer
AI algorithms (especially deep learning) have made great progress in medical image recognition; convolutional variational autoencoders and other methods have numerous applications in this field. In fact, as early as 2001, neural networks were used to analyze endoscopic ultrasound images to distinguish pancreatic cancer from focal pancreatitis. A program was designed that could distinguish pancreatitis from pancreatic cancer by extracting pixel features from images, showing a high accuracy rate of 89%. Given the state of image diagnosis technology at that time, the images were relatively simple, but with the help of computer neural networks, differential diagnosis became easier and achieved higher accuracy. Since then, neural network analysis images have been used in research to differentiate pancreatic cancer from chronic pancreatitis. This method involves collecting image data into a vector form and then converting it into a hue histogram. The sensitivity, specificity, and accuracy of this method in the differential diagnosis of benign and malignant pancreatic lesions were 91.4%, 87.9%, and 89.7%, respectively.
Pancreatic cystic lesions are often considered an important sign of pancreatic cancer. Machine learning is used to extract the imaging features of these cystic lesions, select and classify those features, and then use them to predict benign or malignant pancreatic cystic lesions. In this process, first, image acquisition is conducted uniformly, the edge of the suspected lesion object is delineated, and the three-dimensional (3D) shape of the variant is obtained. Then the features of suspected diseases in the image are extracted including the structure, density, and shape. AI software is used for in-depth learning, the features are screened and analyzed, and the imaging output results are obtained. The obtained results, proteomics, and patient data are entered into the machine learning model as the input layer to generate a predictive model, which can help clinicians in the differential diagnosis of benign and malignant pancreatic cysts.The entire process is shown in Figure 3.
Figure 3 The entire process.
Machine learning is used to extract the imaging features of these cystic lesions, select and classify those features, and then use them to predict benign or malignant pancreatic cystic lesions. In this process, first, image acquisition is conducted uniformly, the edge of the suspected lesion object is delineated, and the three-dimensional shape of the variant is obtained. Then, the features of suspected diseases in the image are extracted, including the structure, density, and shape. AI software is used for in-depth learning, the features are screened and analyzed, and the imaging output results are obtained. The obtained results, proteomics, and patient data are entered into the machine learning model as the input layer to generate a predictive model, which can help clinicians in the differential diagnosis of benign and malignant pancreatic cysts.
Over the past 20 years, with the popularization and development of computed tomography (CT), magnetic resonance imaging, and positron emission tomography-CT, medical staff has been able to obtain more clear imaging data. However, because of human limitations, they cannot achieve zero errors, and diagnostic efficiency is not high. Furthermore, it is time consuming to train professional radiologists. Moreover, the image itself can only reflect the internal structure of the patient at a certain time and from a certain angle; thus, slight changes can be difficult to be detect with the naked eye. As such, reliable AI can improve the accuracy of image diagnosis.
As mentioned above, manual diagnosis has shortcomings such as subjective judgment, a lack of repeatability, and low accuracy. Recent research on using convolutional layer neural networks to recognize CT in pancreatic cancer diagnosis may provide a way to overcome such shortcomings. An AI designed for one related study consisted of two parts: training and verification. First, a patient data database is established, image data are collected, and an image database is established. Then the feature extraction, area generation (RPN), and classification and regression networks are established. In the AI network, the input image is first converted into a convolutional feature graph, and the RPN parameters are adjusted through the feature map to generate the ROI feature vectors. Then the RPN parameters are put into the convolutional layer, and a certain model is used for regression and classification. Next, the regression parameters are generated into new RPN parameters, and the two RPN parameters are updated only for the unique network layer of RPN through machine learning. The RPN parameters are then generated by the regression parameters to fine-tune the unique convolutional layer. Using a reserved verification group input model, the Secure Global Desktop network is trained by back-propagation and random gradient descent, and the network weights and parameters can be constantly updated and optimized. Finally, the final model is obtained as an AI diagnosis system. The receiver operating characteristic curve of the experimental results reached 0.9632. The AI in that study needed only 20 s to identify images and was more objective and effective than traditional diagnosis methods. It was noted, however, that while this method showed high accuracy in the diagnosis of pancreatic cancer, it does not mean AI can replace specialists; rather, it provides an auxiliary tool for diagnosis.
Although AI has good prospects for image diagnosis, it also has limitations, and the process of model training is inseparable from the assistance of artificial diagnosis. In theory, the ultimate goal of diagnostic accuracy is infinitely close to the imaging doctor. Therefore, how to make good use of this to make AI more intelligent may be an important problem to be solved in future research. In fact, the application of AI in imaging has been investigated by experts in many fields, and it also requires knowledge from many fields. Such projects create a platform for imaging experts to communicate with computer experts. The result is that an AI system is established that uses a deep learning algorithm to collect and analyze CT images of the pancreas. The experimental group image data and the normal control image data are imported into the program. Through two matrices and the application of a filter, statistics, texture, shape, and other data are obtained. Then the pancreatic ductal adenocarcinoma and the normal control are distinguished by data processing, statistical analysis, and the random forest model.
The relationship between AI and imaging involves knowledge from various fields such as pathology, radiology, oncology, and computer science. Thus, a more intelligent AI system may be built through the combined work of experts from multiple fields. The AI image acquisition discussed above is based on segmenting the pancreas from the image. The traditional segmentation method is a top-down simulation fitting method based on a large amount of map input and fixed pancreatic label fusion. However, there is also a bottom-up pancreatic segmentation method that subdivides the aggregated image region into a pancreatic region and a nonpancreatic region. The segmentation is based on the visual features of the image itself, which can improve the accuracy of pancreatic segmentation. It has been reported that the bottom-up pancreas segmentation method has been optimized. With the improvement of deep convolutional neural networks, this method can deal with the highly complex appearance of the pancreas in CT images.
Based on the above, we can see that the application of AI in the imaging diagnosis of pancreatic cancer has made considerable advances and is constantly improving.
Application of AI in the pathological diagnosis of pancreatic cancer
Pathologists need to identify diseased tissues in different tissue sections, which is a time-consuming and laborious process. Even experienced professionals may have the risk of subjective judgment. As with the application of AI in imaging diagnosis, AI is also important in the field of pathology, wherein tissue sections are digitized by a computer. First, the AI system divides the lumen and nucleus from tissue fragments and extracts feature vectors from tenfold epithelial nuclei. Different cells have different feature vectors. An epithelial nucleus algorithm is used to identify epithelial nuclei. Then, the morphological features of the diseases that can be diagnosed are extracted. Finally, AI classifiers are used for classification. These classifiers include Bayesian classifiers, k-nearest neighbors, support vector machines, and ANNs. Based on an automatic learning framework, cells can be segmented more accurately by combining bottom-up and top-down information. After collecting patient tissue samples, the tissue photographs are uniformly collected. A convolutional neural network model of a deep convolutional neural network is used to generate a probability map of tissue nuclear distribution. Then the iterative region merging method is used to initialize the shape of the probability graph. Next, combining a sparse shape model with stable selection and a local repulsive deformation model, a new segmentation algorithm is proposed to separate a single nucleus.
A significant advantage of this framework is that it is suitable for different stained histopathological images. Because of the feature-learning characteristics of deep cellular neural networks and the characteristics of high-level shape prior modeling, this proposed method is sufficiently universal and can be applied to different staining specimens and various types of histopathological identification. This model is not only less affected by the overlap of pathological tissues and cells but is also relatively insensitive to image noise and uneven intensity. Different tissue-staining datasets are tested, which can identify and label the concentrated area of the nucleus.
In addition to AI classifiers, neural networks also play an important role in image analysis to determine whether the pathology is benign or malignant. After collecting a certain amount of fine-needle aspiration pathology of the pancreatic tumor, a pathological image is captured for preprocessing (image gray conversion and noise reduction). Then, the K-means clustering algorithm is used to extract the highest value of pixels until all of the pixels are equal. The part of the image that needs to be identified is segmented so the tissue can obtain the basic nuclear features, which can be used to evaluate cellular morphological features. These features are input into the AI multilayer perceptron (a feedforward nonlinear neural network) as input vectors, and the decisions made by this perceptron are sent to the second layer perceptron using image evaluation. Because there are a certain number of validated cases, the diagnostic accuracy of benign and malignant lesions can be evaluated using statistical methods (logistic regression, multiple regression, area under the curve, and R-squared). Different from imaging diagnosis, pathological diagnosis pays more attention to accuracy. Thus, AI has a lot of room for improvement in the accuracy of auxiliary diagnosis, which will inevitably take a long time to develop.