CT classification model of pancreatic serous cystic neoplasms and mucinous cystic neoplasms based on a deep neural network

At present, numerous challenges exist in the diagnosis and treatment of pancreatic SCNs and MCNs, such as the rational selection of imaging evaluation methods, the key points of correct imaging diagnosis, and conservative observation or surgical resection. SCNs are benign lesions that exhibit slow growth and a low malignant rate. Most of these lesions do not require surgery. However, if the tumor growth rate is significantly accelerated (diameter growth rate is greater than 2 cm/year) or signs of invasive growth are noted, surgical treatment should be considered [11]. MCNs have the potential to develop into pancreatic cancer, a recent review of 90 resected MCNs found that 10% of them contained either high-grade dysplasia or pancreatic cancer [12]. Therefore, it is necessary to seek an optimal strategy for the diagnosis of pancreatic SCNs and MCNs. Compared with CT and MRI examination, ultrasound examination of pancreatic lesions is susceptible to gastrointestinal gas, obesity, etc., and at contrast-enhanced ultrasound (CEUS) cystic tumors was correctly diagnosed with an sensitivity of 78.2% and with an NPV (negative predictive value) of 97.1%, which are much lower than that of CT examination [13]. Therefore, CT and MRI evaluation of the type of pancreatic SCNs and MCNs has become an important factor in determining treatment.

In the past, CT diagnosis of pancreatic SCNs and MCNs relied on the subjective empirical judgment of the radiologist [14], and sometimes they could use MR images for comprehensive analysis. Radiomic analysis methods eliminate the need for extensive clinical experience and avoid the interpretation of empirical imaging metrics [15]. However, at present, many radiomic methods rely on a large number of pre-defined features to quantitatively describe the characteristics of medical images, such as tumor volume and texture [16], and statistical methods are used to select the features most relevant to the results. Finally, the machine learning method is used to establish the diagnosis and prediction model. Among them, the commonly used methods include Logistic Regression, Support Vector Machine (SVM), and Random Forest [17]. Computer deep learning is a new field in machine learning research that can be used for medical image classification, segmentation, recognition, and brain function research [8, 9]. For example, DNN can make computers simulate the human brain for analytical learning and the human brain visual mechanism to automatically learn the abstract features of each level of data to better reflect the essential characteristics of the data. Thus, in this study, we used DNN to extract and classify the features of the lesions’ images and obtained the expected effect.

Image segmentation involves the extraction of ROI. This study first relied on an experienced radiologist to manually outline and segment the pancreatic cystic lesions in the original single-channel CT image. It is easy to produce volume effects when lesions are manually outlined inaccurately, causing erroneous information to be incorporated into the area of interest. Therefore, we preprocessed the single-channel manual outline ROI image to construct a Multi-channel image. During the conversion process, the window width, window level, and gradient magnitude are readjusted to enhance the difference in image information between the lesion and the surrounding tissue. Then, the Canny operator is used to obtain edge information of the lesion image, this effective edge detection method is based on the notion that the edge of the tumor is indicated by the local discontinuity of the image, such as sudden changes in grayscale, sudden changes in color, and sudden changes in texture structure [18]. Our study shows that six indicators (sensitivity, specificity, precision, accuracy, F1 score, and AUC) of classification effect using Multi-channel images are all better than using single-channel manual outline ROI image for the classification of pancreatic SCNs and MCNs. The Multi-channel image obtained after human–computer interaction and computer post-processing provides a more accurate lesion area image. Thus, Multi-channel CT images can better distinguish pancreatic SCNs from MCNs than single-channel manually outlined ROI images.

CNN (Convolutional Neural Network) is one type of DNN method [19] that extracts corresponding features in different layers, both AlexNet and ResNet represent excellent CNN methods [20, 21]. In this study, six indicators (sensitivity, specificity, precision, accuracy, F1 score, and AUC) of classification efficacy yielded better classification results with DNN ResNet than with AlexNet when using Multi-channel images and the Softmax classifier for the classification of pancreatic SCNs and MCNs with pathological results serving as the gold standard. The features extracted by ResNet were more representative than those extracted by the currently commonly used feature extraction methods (wavelet, LBP, HOG, GLCM, Gabor) for the classification of pancreatic SCNs and MCNs in the context of a multi-channel image and the same classifier, including the SVM classifier, KNN, and Bayes classifiers. ResNet performed better than AlexNet for the classification of pancreatic SCNs and MCNs. The following questions arise: Will the study draw the wrong conclusion if Softmax classifiers are exclusively used to classify lesions after extracting image features using ResNet and AlexNet? If the classification is performed by other classifiers, will the results change? The answer is no because when the features of different dimensions are entered into the same type of classifier for classification, the extracted features are the determining factor of the classification instead of the classifier. Alex's network structure contains fewer layers than ResNet, and the number of extracted features is far less than obtained with ResNet. Thus, ResNet improves the diagnostic efficacy of pancreatic SCNs and MCNs.

The role of the classifier is to learn the classification rules using the given and known training data categories and then classifying (or predicting) unknown data [22]. MMRF-ResNet uses a random forest classifier to integrate the classification probabilities of the KNN, Bayes, and Softmax classifiers. The random forest classifier is employed in this study because it can provide a certain classifier weight according to the classification result of a certain classifier during the analysis of the training data. Classifiers with better efficacy have higher weights, and classifiers with poorer efficacy have lower weights. The random forest classifier reasonably and comprehensively judges the classification results of the three classifiers and integrates the classification probabilities. Classification results of the majority voting rule method are consistent with the classification results of two or more classifiers among Softmax, KNN, and Bayes classifier. Thus, the results using the random forest classifier method (MMRF-ResNet) are more realistic than the results using the majority voting rule method.

There are still some shortcomings in this study. First, the study is retrospective, the number of samples is not large enough, and there are inherent limitations. Second, MMRF-ResNet requires that the lesion image is manually outlines for segmentation, and automatic segmentation is not possible. Third, we only analyzed the region of interest in images and did not analyze location information of the lesions (such as the head, body, and tail of the pancreas) and patient clinical information, such as gender, age, family history, and clinical symptoms, and the characteristics of the tumor have not been considered: size, grading, vascularization etc., for example are informations that can complete the clinical situation and they could be very useful notions. Fourth, we did not use the patient's MR examination images for comprehensive analysis. Fifth, radiomics features are affected by CT scanner parameters such as reconstruction kernel or section thickness, thus obscuring underlying biologically important texture features. This study did not use compensation methods to correct the variations of radiomic feature values caused by using different CT protocols [23]. Sixth, in the comparative study of single classifiers and multiclassifiers, given that the Softmax classifier uses more sampling points than the other three methods, the obtained AUCs ranked from the largest to the smallest are as follows: Softmax classifier, MMRF-ResNet, KNN classifier, and Bayes classifier. These results suggest that the diagnostic performance of MMRF-ResNet is lower than that of the Softmax classifier, but this result is inaccurate due to the number of sampling points. Finally, In how many cases the diagnosis match imaging findings and computed classification? The above shortcomings have yet to be further addressed.

In conclusion, in this study, Multi-channel CT images were obtained through preprocessing based on single-channel manual outline ROI images, and ResNet was used to extract CT image features of pancreatic SCNs and MCNs. The random forest classifier is used to integrate the classification probabilities of the KNN, Bayesian, and Softmax classifiers to determine the CT image properties of pancreatic SCNs and MCNs. Finally, a better classification result was obtained relative to the commonly used radiomics methods, suggesting that MMRF-ResNet is an ideal CT classification model for distinguishing between pancreatic SCNs and MCNs.

留言 (0)

沒有登入
gif