If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Histogram array and convolutional neural network of DWI for differentiating pancreatic ductal adenocarcinomas from solid pseudopapillary neoplasms and neuroendocrine neoplasms
1 Yan-Jie Shi and Hai-Tao Zhu contributed equally to this article.
Yan-Jie Shi
Footnotes
1 Yan-Jie Shi and Hai-Tao Zhu contributed equally to this article.
Affiliations
Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education), Department of Radiology, Peking University Cancer Hospital & Institute, No.52 Fu Cheng Road, Hai Dian District, Beijing 100142, China
1 Yan-Jie Shi and Hai-Tao Zhu contributed equally to this article.
Hai-Tao Zhu
Footnotes
1 Yan-Jie Shi and Hai-Tao Zhu contributed equally to this article.
Affiliations
Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education), Department of Radiology, Peking University Cancer Hospital & Institute, No.52 Fu Cheng Road, Hai Dian District, Beijing 100142, China
Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education), Department of Radiology, Peking University Cancer Hospital & Institute, No.52 Fu Cheng Road, Hai Dian District, Beijing 100142, China
Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education), Department of Radiology, Peking University Cancer Hospital & Institute, No.52 Fu Cheng Road, Hai Dian District, Beijing 100142, China
Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education), Department of Radiology, Peking University Cancer Hospital & Institute, No.52 Fu Cheng Road, Hai Dian District, Beijing 100142, China
Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education), Department of Radiology, Peking University Cancer Hospital & Institute, No.52 Fu Cheng Road, Hai Dian District, Beijing 100142, China
Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education), Department of Radiology, Peking University Cancer Hospital & Institute, No.52 Fu Cheng Road, Hai Dian District, Beijing 100142, China
CNN is proposed to differentiate PDACs from non-PDACs by converting DWI into an image of histogram array.
•
AUCs of histogram array and CNN for diagnosing PDACs are 0.896, 0.846, and 0.839 in training, validating and testing cohorts.
•
The performance of histogram array and CNN for diagnosing PDACs is higher than that of DWI parameters in all patients.
Abstract
Purpose
This study aimed to investigate the diagnostic performance of the histogram array and convolutional neural network (CNN) based on diffusion-weighted imaging (DWI) with multiple b-values under magnetic resonance imaging (MRI) to distinguish pancreatic ductal adenocarcinomas (PDACs) from solid pseudopapillary neoplasms (SPNs) and pancreatic neuroendocrine neoplasms (PNENs).
Methods
This retrospective study consisted of patients diagnosed with PDACs (n = 132), PNENs (n = 45) and SPNs (n = 54). All patients underwent 3.0-T MRI including DWI with 10 b values. The regions of interest (ROIs) of pancreatic tumor were manually drawn using ITK-SNAP software, which included entire tumor at DWI (b = 1500 s/m2). The histogram array was obtained through the ROIs from multiple b-value data. PyTorch (version 1.11) was used to construct a CNN classifier to categorize the histogram array into PDACs, PNENs or SPNs.
Results
The area under the curves (AUCs) of the histogram array and the CNN model for differentiating PDACs from PNENs and SPNs were 0.896, 0.846, and 0.839 in the training, validation and testing cohorts, respectively. The accuracy, sensitivity and specificity were 90.22%, 96.23%, and 82.05% in the training cohort, 84.78%, 96.15%, and 70.0% in the validation cohort, and 81.72%, 90.57%, and 70.0% in the testing cohort. The performance of CNN with AUC of 0.865 for this differentiation was significantly higher than that of f with AUC = 0.755 (P = 0.0057) and α with AUC = 0.776 (P = 0.0278) in all patients.
Conclusion
The histogram array and CNN based on DWI data with multiple b-values using MRI provided an accurate diagnostic performance to differentiate PDACs from PNENs and SPNs.
Pancreatic ductal adenocarcinomas (PDACs) are the most commonly encountered pancreatic lesions followed by pancreatic neuroendocrine neoplasms (PNENs) and solid pseudopapillary neoplasms (SPNs) [
]. The treatment approach and prognosis of patient are significantly different, accurate diagnosis for theses pancreatic tumors is important in clinical practice. PDAC is a highly aggressive malignant neoplasm with an overall 5-year survival of 9%, and aggressive surgical approaches with extensive lymph node dissections are widely applied [
]. Sometimes, similar appearance of these pancreatic neoplasms may result in the confusing interpretation of computed tomography (CT) or magnetic resonance imaging (MRI) data when the radiological findings are atypical [
]. This confusion may lead to misdiagnosis and subsequent inappropriate treatment strategies. Therefore, timely and accurate imaging diagnosis to differentiate PDACs from PNENs and SPNs before treatment is important because it guides surgery planning and selection of treatment strategy, and helps determine patient prognosis.
Diffusion weighted imaging (DWI) is a noninvasive functional technique in MRI to evaluate pancreatic neoplasms because it provides functional information on the Brownian motion of water molecules, and reflects the underlying tumor microstructure [
]. The apparent diffusion coefficient (ADC) value obtained based on only two b-values is commonly applied for assessing the tumor, but this model cannot reflect the complicated tumor environment [
]. Therefore, several non-Gaussian models, known as the stretched exponential model (SEM), intravoxel incoherent motion (IVIM) model, and diffusion kurtosis imaging (DKI) model, have been proposed for detecting, assessing, and diagnosing pancreatic tumors [
Radiomics analysis based on diffusion kurtosis imaging and T2 weighted imaging for differentiation of pancreatic neuroendocrine tumors from solid pseudopapillary tumors.
Establishment of a multi-parameters MRI model for predicting small lymph nodes metastases ( <10 mm) in patients with resected pancreatic ductal adenocarcinoma.
]. All models use several parameters to assess the neoplasm, but the environment in the neoplasm is more complicated than these ideal models. The histogram array obtained by converting DWI data into a 2D image is a new technique independent of any predefined models, which has been applied to predict pathological complete response after neoadjuvant treatment in rectal cancer [
Convolutional neural networks (CNNs), a class of deep learning techniques, have been applied recently to tackle radiological problems. CNNs can extract the most representative features from CT or MRI, whereas radiologists just focus on radiological findings, such as tumor size, tumor signal in T1WI, T2WI and DWI, or enhanced pattern [
]. Therefore, this methodology can avoid the subjective difference and make them independent of prior human knowledge and achieve a high degree of fault tolerance [
Development and validation of a combined nomogram model based on deep learning contrast-enhanced ultrasound and clinical factors to predict preoperative aggressiveness in pancreatic neuroendocrine neoplasms.
]. The CNN models based on imaging have been shown to be effective and powerful for detecting pancreatic tumors, predicting pathological grading of PNENs, and differentiating between malignant and benign pancreatic cystic lesions [
Deep learning for World Health Organization grades of pancreatic neuroendocrine tumors on contrast-enhanced magnetic resonance images: a preliminary study.
Preoperative prediction of pancreatic neuroendocrine neoplasms grading based on enhanced computed tomography imaging: validation of deep learning with a convolutional neural network.
We hypothesized that histogram array and CNN based on DWI data might be helpful in differentiating PDACs from PNENs and SPNs. Therefore, we aimed to implement the histogram array and the CNN model to distinguish PDACs from PNENs and SPNs using DWI with multiple b values. The performance of the CNN model for this differentiation was thoroughly evaluated.
2. Materials and methods
This retrospective study was approved by our institutional review board and ethical committee. The requirement for informed consent was waived.
2.1 Study participants
A total of 231 patients were included in this study. The inclusion criteria were as follow: (1) Patients underwent pancreatic 3-Tesla MRI examination including DWI before surgery; (2) patients received pancreatic surgery and were pathologically confirmed for PDACs, PNENs or SPNs; and (3) patients did not receive any treatment, such as chemotherapy or radiotherapy, before surgery. The exclusion criteria were as follow: (1) The DWI image quality was inadequate for analysis due to noise or artifacts; (2) DWI data was missing; (3) the scanning parameters of DWI were variant. Finally, 132 patients with PDACs diagnosed between January 2015 and June 2019, and 45 with PNENs and 54 with SPNs diagnosed between January 2011 and December 2019 in our hospital were enrolled in this study. The patients were randomly allocated to a training cohort with 92 patients, a validation cohort with 46 patients and a testing cohort with 93 patients. Fig. 1 shows the complete patient enrollment process.
All patients were scanned using a 3-Tesla MRI scanner device (Discovery MR 750; GE Healthcare, WI, USA) equipped with an eight-channel phased array body coil in the supine position. The MRI scheme included the following sequences: axial T1WI and T2WI, coronal T2WI, axial DWI and contrast-enhanced MRI using lava flex sequence. DWI was axially obtained using the single-shot echo-planar imaging sequence with a time of repetition of 7000 ms and time of echo of 60 ms. In DWI, the field of view was 340 mm; the matrix was 128 × 128; slice thickness/slice gap was 3.0/0.5 mm. The multiple b-values were 0, 20, 50, 100, 200, 600, 800, 1000, 1200 and 1500 s/mm2. The numbers of excitation (NEX) of b values = 0–200, 400–1000, 1200 and 1500 s/mm2 were 1, 2, 4, and 6, respectively. The Integrated Parallel Acquisition Technique imaging option was applied with a factor of 3. Distortion correction was also used to avoid or reduce artifacts. The fat suppression technique was used in DWI to reduce the chemical shift artifacts.
2.3 Tumor delineation
The size, location, and shape of the pancreatic lesions, found using MRI, were correlated with the surgical and pathological findings by the radiologists. The radiologists confirmed that each pancreatic lesion in MRI was in accordance with the surgical and pathological results when compared the aforementioned features with MRI and pathology. The region of interest (ROI) of entire pancreatic tumor was manually drawn using the ITK-SNAP software (version 2.2.0; www.itksnap.org). The tumor was delineated on each slice of DWI at a b-value of 1500 s/mm2 by two radiologists (Y-L liu and Y-Y Wei, 8 and 6 years of abdominal MRI experience, respectively) who were blinded to the pathological results. The two radiologists independently drew the ROIs of pancreatic tumors. The tumor often showed high signal intensity at a b value of 1500 s/mm2 of DWI. When drawing the pancreatic tumors on DWI, the radiologists contained the entire tumor according to other sequences. When the tumor did not show high signal on DWI compared with the normal pancreatic tissue, the ROIs were determined by using other MRI sequences, such as T1WI, T2WI and contrast imaging. In the case of multiple lesions in the pancreas, a lesion with the highest probability of malignancy was chosen by observers.
2.4 Histogram array performance
The DWI images were converted into a histogram array according to the following steps:
1.
The DWI signal S (x) of each pixel was converted into its logarithmic value log [S(x)].
2.
The histogram was calculated from the voxels inside ROI for each b-value. The number of gray-level bins was set to 70.
3.
The histograms vertically concatenated in an ascending order of b-values to construct an image.
Finally, this converted image was input into the CNN model. Fig. 2 presents the converted image from 10 histograms.
Fig. 2The conversion of DWI images into a histogram array of a 2D image.
Data augmentation was performed to enlarge the sample size by randomly selecting half of all the voxels in each ROI to generate a virtual subject. Subjects in different categories were augmented at different times to approach an equal sample size for each category. In this study, the augmentation time for PDACs, PNENs and SPNs was 34, 100 and 86 respectively in the training cohort and 35, 100 and 82 times respectively in the validation cohort. No augmentation was performed in the testing cohort.
2.6 Implementing CNN for the differentiation of PDACs from PNENs and SPNs
The CNN model was implemented using PyTorch (version 1.11) based on Python (version 3.7). The CNN model consisted of two convolutional layers, two max-pooling layers, two rectified linear unit (ReLU) layers and two linear (fully connected) layers. The label was the one-hot format of the pathological ground truth of PDACs, PNENs, or SPNs. Cross-entropy was used as the loss function. Stochastic gradient descent was used for optimization. The output of the network was the probability of each category. The network structure was plotted as shown in Fig. 3. The learning rate was set to 10−4, and the batch size was set to 50. The highest accuracy on the validation cohort appeared at the 2862nd epoch. The corresponding network was selected as the final model to be evaluated on the testing cohort.
Fig. 3CNN model for classifying the histogram array image into PDAC, PNEN, and SPN.
Non-Gaussian DWI models were voxel-wise fitted in the ROI after applying a 3 × 3 smoothing kernel to each slice. D, D*, and f were extracted from the IVIM model [
Evaluation of hepatic focal lesions using diffusion-weighted MR imaging: comparison of apparent diffusion coefficient and intravoxel incoherent motion-derived parameters.
]. D is the slow component of diffusion representing pure molecular diffusion; D* is the fast component of diffusion reflecting perfusion; and f is the volume fraction of the protons quantifying the intravascular component in the microcirculation. Distributed diffusion coefficient (DDC) and α were obtained from SEM. DDC is the mean intravoxel diffusion rate, and α is the intravoxel water molecular diffusion heterogeneity index [
]. Mean kurtosis (MK) and mean diffusivity (MD) were calculated from the DKI model. MK measures the degree of deviation from Gaussian distribution, and MD is the diffusion coefficient after non-Gaussian correction [
Diffusion and perfusion MR parameters to assess preoperative short-course radiotherapy response in locally advanced rectal cancer: a comparative explorative study among standardized index of shape by DCE-MRI, intravoxel incoherent motion- and diffusion kurtosis imaging-derived parameters.
SPSS Statistics version 22.0 (IBM Corp., NY, USA) was used for the statistical analysis of patient characteristics. Continuous variables, such as age, were expressed as mean ± standard deviation and were compared using the Students t-test. Categorical variables, such as sex, were compared using the chi-square test. The receiver operating characteristic (ROC) curves were plotted using MATLAB software (R2017b; MathWorks, MA, USA). The area under the curve (AUC) was compared using the DeLong test. The threshold was determined by maximizing the sum of sensitivity and specificity. The corresponding positive predictive value (PPV) and negative predictive value (NPV) were also calculated. The intraclass correlation coefficient (ICC) was calculated to evaluate the interobserver agreement on ROI delineation. A coefficient larger than 0.80 was considered as almost perfect agreement. Post hoc power analysis was conducted in the testing cohort, taking the null hypothesis of AUC 0.70 and the AUC of the constructed CNN model as the alternative hypothesis using a two-sided z-test at a significance level of 0.05. A value of P < 0.05 (two-tailed test) indicated a statistically significant difference.
3. Results
3.1 Patient characteristics
The characteristics of all patients in this study are depicted in Table 1, which indicates no significant differences in age, sex and pathological diagnosis among the training, validation and testing cohorts (all P > 0.05).
Table 1Demographic and pathological characteristics of patients in the training, validation and testing cohorts.
Two radiologists who independently drew the ROIs of pancreatic tumors, achieved perfect agreement with the ICC value of 0.810. The number of patients with PDACs, PNENs, and SPNs were 53, 18 and 21 in the training group, and 26, 9, and 11 in the validation group, respectively. After augmentation, the number of patients with PDACs, PNENs and SPNs were 1855, 1818 and 1827 in the training group, and 936, 909, and 913 in the validation group, respectively.
3.3 Performance of histogram array and CNN in distinguishing PDACs from SPNs and PNENs
The CNN performance in distinguishing PDACs from non-PDACs is summarized in Table 2, Table 3. All the statistical results excluded data augmentation. The CNN model produced a score for each pancreatic tumor. A score larger than the cutoff value of 1.49 suggested the diagnosis of PDAC. The overall diagnostic accuracy for this differentiation was 90.22%, 84.78% and 81.72% in the training, validation, and testing groups, respectively. The sensitivity was 96.23%, 96.15%, and 90.57% in the training, validation and testing groups. The specificity was 82.05%, 70.0% and 70.0% in the training, validation, and testing groups. The AUCs of CNN for this differentiation were 0.896, 0.846, and 0.839 in the training, validation, and testing groups, respectively. The post hoc power was 0.73 for the testing group.
Table 2AUCs of CNN and parameters of DWI models for diagnosing PDACs and non-PDACs.
Models
Training cohort (n = 92)
Validation cohort (n = 46)
Testing cohort (n = 93)
CNN
0.896 (0.816–0.975)
0.846 (0.719–0.973)
0.839 (0.752–0.926)
D
0.663 (0.548–0.778)
0.783 (0.638–0.927)
0.720 (0.606–0.834)
D*
0.513 (0.393–0.633)
0.698 (0.546–0.850)
0.555 (0.435–0.674)
f
0.712 (0.606–0.818)
0.890 (0.792–0.988)
0.719 (0.611–0.827)
DDC
0.603 (0.479–0.727)
0.710 (0.552–0.868)
0.534 (0.408–0.660)
α
0.777 (0.678–0.876)
0.854 (0.728–0.979)
0.722 (0.616–0.828)
MD
0.603 (0.481–0.725)
0.731 (0.580–0.882)
0.575 (0.453–0.696)
MK
0.715(0.608–0.822)
0.694 (0.533–0.856)
0.736 (0.626–0.846)
AUCs: area under curves; CNN: convolutional neural network; D: diffusion; D*: perfusion; α: diffusion heterogeneity index; DDC: distributed diffusion coefficient; f: fraction; MD: mean diffusivity; MK: mean kurtosis.
The data in parentheses are 95 % confidence interval (CI).
3.4 Comparing with CNN and parameters of DWI in differentiating PDACs from SPNs and PNENs
Table 2, Table 3, and Fig. 4 show the comparison between the CNN model and DWI parameters of the IVIM (D, D* and f), DKI (MD and MK) and SEM (DDC and α) models. The α value with an AUC value of 0.776 indicated the best performance, followed by f (AUC = 0.755) and MK (AUC = 0.721) in all patients. The α value of PDACs with value of 0.70 ± 0.13 was higher than that of SPNs and PNENs with value of 0.60 ± 0.11. A value of α >0.69 indicated the diagnosis of PDAC. The f value of PDACs with 0.37 ± 0.12 was lower than that of SPNs and PNENs with 0.47 ± 0.11. An f value <0.37 indicated PDAC. Fig. 4 compares ROC curves with CNN, α and f values for distinguishing PDACs from non-PDACs. The CNN performance with an AUC of 0.865 for this differentiation was significantly higher than that of f (Z = 2.762 and P = 0.0057) and α (Z = 2.200 and P = 0.0278) in all patients. However, no significant difference was observed between the CNN model and DWI parameters in the training, validation and testing groups respectively (all P values > 0.05).
Fig. 4ROC curve of the histogram array and CNN model, α, and f for distinguishing PDACs from PNENs and SPNs in all patients (A), training cohort (B), validation cohort (C) and testing cohort (D). AUC values of the histogram array and CNN model, α, and f for differentiating PDACs from SPNs and PNENs were 0.865, 0.776, and 0.755 in all patients (A), 0.896, 0.777, and 0.712 in the training cohort (B), 0.846, 0.854, and 0.890 in the validation cohort (C), and 0.839, 0.722, and 0.719 in the testing cohort (D).
DWI is an essential diagnostic tool for evaluating pathological conditions in the pancreas and is increasingly applied in routine clinical practice. However, DWI assessment might be limited in evaluating pancreatic lesions. Fukukura et al. found that DWI was not useful for delineating 47% of PDACs because of the hyperintensity of the pancreatic parenchyma distal to cancer [
]. Further, DWI parameters might be affected by many complicated and potential factors including DWI techniques, magnetic field inhomogeneity, field strength, image noise and artifacts. Sometimes these factors might potentially change the DWI parameters [
]. Ha et al. found that although all available studies demonstrated the difference in ADC value between PDAC and auto immune pancreatitis, some studies showed a lower ADC value in the PDAC group, while others showed the opposite result [
]. Although it was possible to use SEM, IVIM or DKI models to evaluate pancreas, all these methods used predefined mathematic models to approximate the decay curve of DWI signal. Therefore, we proposed a deep learning method to analyze the DWI data independent of any DWI models. It might avoid errors induced by inappropriate approximation.
Radiomics and deep learning are commonly used in the classification task for radiological images such as CT and MRI. Radiomics generally uses traditional classifiers that requires predefined features. So, the histogram array is not suitable to apply in radiomics. Deep learning uses CNN to automatically extract features from the images without predefinition. CNN has been applied to diagnose and differentiate pancreatic lesions and demonstrated a high-level performance. Ma et al. found that CNN based on CT could be suitable for pancreatic cancer detection and showed an excellent performance compared with radiologists' manual interpretations [
]. Gao et al. showed the potential of the CNN model for predicting the World Health Organization–classified PNENs with AUC values of 0.92 and 0.88 in the training and validation groups on contrast-enhanced MRI [
Deep learning for World Health Organization grades of pancreatic neuroendocrine tumors on contrast-enhanced magnetic resonance images: a preliminary study.
]. Uguon et al. showed that the CNN model achieved high performance of differentiation between mucinous and serous cystic neoplasms with an AUC value of 0.88 [
]. In this study, CNN based on multiple-b values of DWI demonstrated a level of competence comparable with the parameters of DWI for distinguishing PDACs from PNENs and SPNs. DWI data from multiple b values constructs a 4-dimensional (4D) dataset by adding a b-value dimension into the 3 spatial dimensions. However, it is quite difficult to directly apply a 4D CNN on a 4D dataset, because it has too many parameters to optimize and require very large computing power. Most current studies use a 2D dataset and a 2D CNN model for classification. It is easy to obtain abundant 2D images for training and it also requires less computing time to train a 2D CNN model.
In our study, DWI images from 10 b values were converted into a histogram array as a single 2D image with a gray level dimension and a b-value dimension. A 2D CNN containing 2 convolution layers were used to classify the histogram array into 3 categories: PDACs, PNENs and SPNs. The simple CNN structure makes it easy to be optimized by training. In addition, the conversion into a histogram array provided a simple and effective method of augmentation by randomly selecting half of all the voxels in each ROI to generate a virtual subject. Consequently, the number of samples was significantly increased in the training cohort. Furthermore, the augmentation in this study increased the number of samples of PNENs and SPNs and balanced the number of sample of PDACs, PNENs and SPNs, which avoided classifying an unlabeled sample in the category due to imbalanced sample in the training cohort. The other advantage of this study was that it combined the CNN and multiple b-values of DWI for differentiating PDACs from SPNs and PNENs. Unlike diffusion and perfusion, the CNN based on DWI might obtain the different information between PDACs and non-PDACs.
In our previous study, we developed two DWI models: PDACs versus PNENs, and PDACs versus SPNs [
Non-gaussian models of 3-tesla diffusion-weighted MRI for the differentiation of pancreatic ductal adenocarcinomas from neuroendocrine tumors and solid pseudopapillary neoplasms.
]. Compared with the DWI models, we developed a CNN model in this study to differentiate PDACs from non-PDACs (PNENs and SPNs). The CNN model with high performance for this differentiation could avoid subjective diagnosis and decrease biopsy rate. This study was not designed to differentiate between PNENs and SPNs because of a similar treatment strategy and favorable prognosis for PNENs and SPNs compared with PDACs. When the diagnosis of non-PDACs was made following the CNN model in this study, the radiomics signature based on DWI and T2WI was recommended to distinguish PNENs from SPNs [
Radiomics analysis based on diffusion kurtosis imaging and T2 weighted imaging for differentiation of pancreatic neuroendocrine tumors from solid pseudopapillary tumors.
This study had several limitations. First, the patients might have a potential selection bias because of the retrospective and single center nature of this study. Second, the number of PNENs and SPNs was relatively small for training CNN models because of the rarity of these tumors. Third, some of the tumor characteristics might not be reflected in our CNN model. Fourth, the CNN model combining DWI and other MRI sequences could not be investigated in this study due to the design of the study involving clinical practice. Fifth, the utility of our CNN model was not assessed in clinical practice; other pancreatic tumors, such as pancreatic cystic tumor, intraductal papillary mucinous neoplasm, and metastases, were not included in this study.
5. Conclusion
CNN is proposed to differentiate PDACs from PNENs and SPNs by converting multiple b values of DWI into an image of a histogram array. The result showed that the CNN model based on DWI was a promising method for diagnosing pancreatic tumors and could assist clinicians in deciding individualized treatment regimens for treating pancreatic neoplasms.
Funding sources
This study was supported by the Beijing Natural Science Foundation (Z200015), the Beijing Municipal Administration of Hospitals Incubating Program (code PX2020046 and PX2018041), the Beijing Hospitals Authority Ascent Plan (Code 20191103), and PKU-Baidu Fund (Grant No. 2020BD027).
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work report in this article.
References
He M.
Liu Z.
Lin Y.
Wan J.
Li J.
Xu K.
et al.
Differentiation of atypical non-functional pancreatic neuroendocrine tumor and pancreatic ductal adenocarcinoma using CT based radiomics.
Radiomics analysis based on diffusion kurtosis imaging and T2 weighted imaging for differentiation of pancreatic neuroendocrine tumors from solid pseudopapillary tumors.
Establishment of a multi-parameters MRI model for predicting small lymph nodes metastases ( <10 mm) in patients with resected pancreatic ductal adenocarcinoma.
Development and validation of a combined nomogram model based on deep learning contrast-enhanced ultrasound and clinical factors to predict preoperative aggressiveness in pancreatic neuroendocrine neoplasms.
Deep learning for World Health Organization grades of pancreatic neuroendocrine tumors on contrast-enhanced magnetic resonance images: a preliminary study.
Preoperative prediction of pancreatic neuroendocrine neoplasms grading based on enhanced computed tomography imaging: validation of deep learning with a convolutional neural network.
Evaluation of hepatic focal lesions using diffusion-weighted MR imaging: comparison of apparent diffusion coefficient and intravoxel incoherent motion-derived parameters.
Diffusion and perfusion MR parameters to assess preoperative short-course radiotherapy response in locally advanced rectal cancer: a comparative explorative study among standardized index of shape by DCE-MRI, intravoxel incoherent motion- and diffusion kurtosis imaging-derived parameters.
Non-gaussian models of 3-tesla diffusion-weighted MRI for the differentiation of pancreatic ductal adenocarcinomas from neuroendocrine tumors and solid pseudopapillary neoplasms.
00012-8&id=gr1.jpg){kind=link}
00012-8&id=gr2.jpg){kind=link}
00012-8&id=gr3.jpg){kind=link}
00012-8&id=gr4.jpg){kind=link}