Applying a radiomics-based CAD scheme to classify between malignant and benign pancreatic tumors using CT images

Abstract

BACKGROUND:

Pancreatic cancer is one of the most aggressive cancers with approximate 10% five-year survival rate. To reduce mortality rate, accurate detection and diagnose of suspicious pancreatic tumors at an early stage plays an important role.

OBJECTIVE:

To develop and test a new radiomics-based computer-aided diagnosis (CAD) scheme of computed tomography (CT) images to detect and classify suspicious pancreatic tumors.

METHODS:

A retrospective dataset consisting of 77 patients who had suspicious pancreatic tumors detected on CT images was assembled in which 33 tumors are malignant. A CAD scheme was developed using the following 5 steps namely, (1) apply an image pre-processing algorithm to filter and reduce image noise, (2) use a deep learning model to detect and segment pancreas region, (3) apply a modified region growing algorithm to segment tumor region, (4) compute and select optimal radiomics features, and (5) train and test a support vector machine (SVM) model to classify the detected pancreatic tumor using a leave-one-case-out cross-validation method.

RESULTS:

By using the area under receiver operating characteristic (ROC) curve (AUC) as an evaluation index, SVM model yields AUC = 0.750 with 95% confidence interval [0.624, 0.885] to classify pancreatic tumors.

CONCLUSIONS:

Study results indicate that radiomics features computed from CT images contain useful information associated with risk of tumor malignancy. This study also built a foundation to support further effort to develop and optimize CAD schemes with more advanced image processing and machine learning methods to more accurately and robustly detect and classify pancreatic tumors in future.

Keywords

Pancreatic cancer detection classification of pancreatic tumor computer-aided diagnosis (CAD)support vector machine image segmentation using deep learning model

1 Introduction

Pancreatic cancer is the seventh leading cause of cancer deaths globally. It was the 11th most common cancer in women and the 12th most common in men. Based on the latest published cancer statistics data [1], in 2021, the estimated new pancreatic cancer cases are approximately 60,430 and deaths from pancreatic cancer are approximately 48,220 in the United States. Although pancreatic cancer accounts for only 3.1 percent of new cancer cases detected, it accounts for approximately 7 percent of deaths in all cancer patients each year. Thus, early detection and diagnosis of pancreatic carcinoma is important for effective management and treatment of this disease. For this purpose, many imaging modalities or techniques including computed tomography (CT), magnetic resonance imaging (MRI) and ultrasound (US) imaging have been traditionally used for detection and diagnosis of pancreatic carcinoma [2].

In current clinical practice, CT is the most cost-effective and extensively used imaging modality or examination method for diagnosis and staging of pancreatic cancer [3]. Pancreatic adenocarcinoma usually presents as an area of low attenuation on enhanced CT images. The spiral CT has been reported to show high sensitivity for detecting suspicious pancreatic tumors, which thus improve the early diagnosis and accurate staging of pancreatic cancer. However, diagnosis of pancreatic cancer by reading and interpreting CT images remains a quite difficult or challenged task for the radiologists in current clinical practice due to the high heterogeneity among the suspicious pancreatic tumors depicting on CT images. It is important to note that some pancreatic adenocarcinomas are isoattenuated and pancreatitis with pancreatic adenocarcinoma may sometimes lead to an overestimation of staging. In order to assist radiologists more accurately and efficiently detect and diagnose pancreatic cancer in reading CT images, developing automated or semi-automated computer software or computer-aided detection and diagnosis (CAD) schemes to help process and analyze CT images has recently been attracting broad research interest among many research groups worldwide.

Since manual dissection of the clinical pancreas and/or pancreatic tumors requires a lot of time and energy from the radiologists and it is also a quite tedious and error-prone task with potentially high inter-reader variability, a number of researches have conducted studies to develop and apply the automated computer-aided schemes in order to more efficiently and accurately segment pancreas depicting on CT images. Since pancreas is characterized by localization uncertainty and morphologic variability, many previously experimental studies indicated that the Dice similarity coefficient (DSC) of pancreatic segmentation results obtained using CT images only reached approximately 71% [4]. Generally, pancreatic segmentation methods can be categorized into top-down methods and bottom-up methods. In top-down methods, earlier information is produced and joined into the one-stage system. There are more top-down methods based on atlas [5], active shape model (ASM) [6] and level set algorithm [7]. However, due to the irregular shape of the pancreas, applying these methods is often difficult to achieve high precision in pancreas segmentation results. For non-Gaussian pathological organs such as pancreas, the bottom-up method has higher accuracy. Recently, to further improve tumor segmentation accuracy, the modified deep learning models have been investigated and tested to segment medical images [8, 9] including pancreas and/or pancreatic tumors [10].

In addition to the development of automated schemes to segment pancreas or tumors, researchers have also developed and tested other CAD schemes aiming to help diagnose or stage pancreatic tumors or cancer using the multi-feature fusion-based machine learning classifiers. For example, one study [11] used Fisher scores to select image features extracted from CT images and then trained a support vector machine (SVM) to classify between the normal and pancreatic cancer cases. The study used a dataset of 23 patients and reported 93.2% classification accuracy. The second study [12] used the least absolute shrinkage and selection operator (LASSO) to select image features and trained an ensemble learning-support vector machine (EL-SVM) to diagnose pancreatic cancer. This study reported 86.6% classification accuracy in classifying 168 CT images with normal and cancer cases. The third study [13] developed and tested a radiomics logistic regression model to predict lymph node metastasis in pancreatic cancer patients. Applying to a dataset of 85 patients, the study reported a prediction accuracy with an area under ROC curve of AUC = 0.841. However, in these CAD studies, pancreatic tumor region in each case was manually segmented by radiologists on one selected CT image slice that is considered depicting with the maximum size of the pancreatic tumor. As a result, these studies had limitations including the tedious or time-consuming human intervention, inter-reader variability in tumor segmentation and only 2D image features can be computed from one selected CT image slice.

In order to address and help overcome some of the limitations or research gaps in developing CAD schemes of pancreatic cancer using CT images in previous studies, we in this study propose to develop and test a new and novel radiomics-based CAD scheme that enables to integrate both automated pancreatic tumor segmentation and machine learning classification tasks into one software package. To build this new CAD scheme, we investigate and take following 5 steps of image preprocessing namely, pancreas segmentation, tumor segmentation, radiomics feature computation and selection, and building a multi-feature fusion-based machine learning classifier to help detect and classify suspicious pancreatic tumors. The details of our study design, experimental procedures and data analysis results are reported and discussed in the following sections of this article.

2 Materials and methods

In this study, we first assembled a retrospective CT image dataset that involves 77 de-identified study cases of patients previously diagnosed with suspicious pancreatic tumors. Among them, CT images of 33 study cases depict the pathology-confirmed pancreas adenocarcinoma and the rest of CT images of 44 cases depict the confirmed benign tumors including neuroendocrine, serous adenoma, necrotizing pancreatitis, interstitial pancreatitis, etc. Based on current clinical guidelines of the response evaluation criteria in solid tumors (RECIST) [14], radiologists have annotated and marked two maximum perpendicular diameters of the central tumor region of each pancreatic tumor in one selected CT image slice for each study case. Figure 1 shows 4 example CT images with the annotated tumor regions selected from 4 study cases.

Fig. 1

Examples of CT images annotated by a radiologist. Based on RECIST guideline, two maximum perpendicular diameters of the central tumor region are measured and marked (as shown on images). In these 4 images, two tumors (on the top) are verified pancreas adenocarcinoma, while two tumors (on the bottom) are intraductal papillary mucinous neoplasm (left) and serous cystadenomas (right), which currently are benign, but has risk to progress to pancreatic cancer.

Then, using these annotated CT images and the pathology-verified tumor diagnostic results as ground-truth, we developed and tested a new CAD scheme that aims to detect and classify between the malignant and benign pancreatic tumors based on one 2D CT image slice previously selected and annotated by the radiologists in each study case. Figure 2 illustrates a flow chat of the proposed new CAD scheme in this study. Following are the brief discussions of each step involved in this CAD scheme.

Fig. 2

A flow chat of a new CAD scheme aiming to classify pancreatic tumors using CT images.

2.1 Image pre-processing

Currently, many of deep learning models [8 –10] have been developed to segment regions of interest (ROIs) depicting on medical images and their performance are commonly assessed based on different open image datasets, which is principally used to analyze and assess the most recent growth segmentation methods. Because of the detriments of dark image lopsidedness, enormous difference changes and commotion, clinical CT images should be pre-processing to increase reproducibility of automated image segmentation schemes and following image feature computation. For the CT images of our dataset, the underlying voxel values of Hounsfield Unit (HU) range from –1024 to 3071. We apply a standardized preprocessing algorithm to map the dark levels of images with all voxels within the range from –100 to 500, and afterward standardize to the range from 0 to 1, which fits the standardization strategy normally utilized in the deep learning models. Thus, the images normalized to the range within [0, 1] are used by deep learning model to conduct pancreas segmentation from CT images in this study.

Next, we also preprocessed CT images using image processing filters to reduce image noise, which may have negative impact on accuracy of image segmentation and the computed image features. Specifically, we first applied a median filter to process each CT images to remove potential impulse noise followed by an edge preserving smoothing using Gaussian filter to enhance tumor boundaries. In this way, the signal-to-noise ratio of the CT images can be improved and/or normalized for images acquired from the diverse clinical practice.

2.2 Pancreas segmentation

After image preprocessing, we utilize an adjusted applicant region technique that depends on a region proposal network (RPN). Using this method enables to create one bouncing box that contains the organ of the ROI namely, one consistent pancreas construction or some spasmodic pancreas structures. After candidate region segmentation, we used a convolutional neural network (CNN) model to conduct the final pancreas segmentation task (as shown in Fig. 3).

Fig. 3

Pancreas segmentation network based on CNN.

In this CNN model to segment pancreas region, we utilized the closest addition to get the last competitor district with the size (212×148) and then used a dilated convolution in which the dilation rate is set to 1. Accordingly, we utilized the dilated convolutions that have the dilated rate of {3, 6}. The state-of-the-art CNNs regularly utilize little convolution bits (3×3) to keep both calculation and number of boundaries included. We further improve the segmentation accuracy by using the LSTM module to simulate the spatial correlation of the surrounding structures with the pancreas. We also picked a powerful system to control the ideal felt field (2 broadened deconvolution layers) and work on the spatial resolution (2 deconvolution layers). As a result, in our deep learning model we only utilized 2 de-convolution layers to assist with recuperating highlight maps at the first CT image resolution rapidly.

To improve and evaluate the accuracy of this pancreas segmentation network, we used a public pancreas CT image dataset downloaded from The Cancer Image Archive (TCIA) of National Cancer Institute of USA for training and testing. The image dataset consists of 82 abdominal elevation 3D CT scans (approximately 70 s after portal contrast) obtained from 53 men and 27 women [15]. Among them, 17 cases were taken from healthy kidneys prior to nephrectomy, and the remaining 65 were selected by radiologists from patients without major abdominal pathology or pancreatic cancerous lesions. Patients’ ages range from 18 to 76 years old with the mean age of 46.8±16.7 years old. CT scans with image resolution of 512×512 pixels, variation in pixel size and slice thickness within 1.5 to 2.5 mm, were acquired using a Philips or a Siemens MDCT machines in which the X-ray tube voltage is 120 kVp [16].

In order to train and optimize the CNN model to segment pancreas regions depicting on CT images, a 3-fold cross-validation method is used to evaluate the performance of pancreas segmentation. Specifically, we randomly selected 55 cases as the training set and the remaining 27 cases as the test set. Through the experiment, the optimal network training parameters were determined and implemented as shown in Table 1. As a result, Figure 4 shows an example of segmentation result by applying this modified CNN-based deep learning model to segment pancreas depicting on one CT image.

Table 1

Parameters used in the CNN-based pancreas segmentation network

Batch Size	Training Epochs	Adam Optimizer	RMSProp Optimizer
4	160	β1 = 0.9	Decay = 0.95
		β2 = 0.99	Momentum = 0.9
		ɛ = e^-6	ɛ = e^-8

Fig. 4

Example of pancreas segmentation in which the segmented pancreas region is pointed by a blue arrow and cued using yellow color.

2.3 Tumor segmentation

After segmenting pancreas, a boundary is created to guide segmentation of pancreatic tumor. We applied a modified region growth algorithm with 6 connected neighborhoods to segment pancreatic tumor region from the segmented pancreas region as described in above step. In developing this tumor segmentation algorithm, the mean intensity values were first calculated and estimated from the lesion areas previously annotated by radiologists. Next, by implementing a random walker algorithm [17] to refine the rough segmentation, the high-quality segmentation can be obtained. Then, a morphological erosion algorithm was used to eliminate the leakage of the image. For hole smoothing and filing, we performed a 3×3 median filtering at last. Figure 5 demonstrates an example of segmented pancreatic tumor region in one selected CT image slice.

Fig. 5

Example of a pancreatic tumor segmentation in one CT image slice in which tumor region are cued in red color.

2.4 Feature Computation and Selection

After pancreatic tumor segmentation, we used a publicly available radiomics software package (MaZda, https://www.eletel.p.lodz.pl/programy/mazda/ [18]) to compute a set of radiomics features, which includes total 1,267 statistics and texture features computed from the segmented tumor region. Then, in order to eliminate clinically irrelevant and redundant image features in the initially computed feature pool, a feature selection process is performed by using following three feature selection criteria in a sequential order namely, entropy analysis, receiving operating characteristics (ROC) analysis and statistical t-test. Specifically, in these criteria, Entropy calculates distance or divergence [19], ROC is used to generate the empirical receiver operating characteristic (ROC) curve and the slope of the random classifier [20], and the T-test calculates absolute value of the two-sample T-test and combined variance estimate [21].

Based on the Pearson correlation coefficient of the extracted features, most of the texture objects have either independent or weak correlation. There are also a few features that show a strong positive correlation and a strong negative correlation. In order to lessen the effect of over-fitting, we evaluated several feature selection techniques, such as Random Forest (RF), Extreme Gradient Boosting (Xgboost), and least absolute shrinkage and selection operator (LASSO). In different comparisons, each selection method has selected different radioactivity characteristics. Regarding feature selection, LASSO chooses the most radiation and Xgboost chooses the least. Table 2 shows the feature selection results of different feature selection methods. In this study, we subsequently chose the results of using LAASO method because the advantage of LASSO is that it outperforms traditional automatic variable selection methods, such as forward, backward, and stepwise, which can produce erroneous results [22]. Due to the small dataset used in this study, we set up a limit or a threshold to select 12 top features extracted by LASSO from the original pool of 1,267 features as shown the last column of Table 2.

Table 2
Feature selection results of different feature selection methods

RF Xgboost LASSO

Height Mean Entropy

Entropy Contrast Gray entropy

Mean Mean and STD of energy Gray variance

Run-length nonuniformity Width Run-length nonuniformity

Gray variance Rectangularity Mean

Rectangularity Elongation Contrast

Area Mean and STD of energy

Width Height

Width

Rectangularity

Run percentage

Complexity

RF	Xgboost	LASSO
Height	Mean	Entropy
Entropy	Contrast	Gray entropy
Mean	Mean and STD of energy	Gray variance
Run-length nonuniformity	Width	Run-length nonuniformity
Gray variance	Rectangularity	Mean
Rectangularity	Elongation	Contrast
Area		Mean and STD of energy
Width		Height
		Width
		Rectangularity
		Run percentage
		Complexity

^*STD – standard deviation.

2.5 Machine learning classifier

After creating the optimal image feature vector of 12 radiomics features, we built a multi-feature fusion-based machine learning model or classifier aiming to distinguish or classify between the malignant and benign pancreatic tumors depicting on CT images. Although many different machine learning classifiers have been investigated and used in developing CAD schemes of medical images, we selected to build a widely used support vector machine (SVM) classifier due to its advantages working with the small image datasets and higher generalizability. SVM is also an efficient hyperplane defining and learning algorithm, which can be easily trained and implemented. Specifically, we selected to use a multi-layer perceptron (MLP) kernel that is generated from the neural organized hypothesis to build the SVM classifier. Because the SVM model using sigmoid kernel function is equivalent to two-layer perceptron neural network, we analyzed that MLP not only maps an eigenvector from the original d-dimensional space, but also maps an eigenvector from an intermediate implicit Hilbert feature space in which the inner product is computed. The learning kernel replaces the regular inward item between the weight vector and the input vector. In this way, the generalization ability of the general function approximator can be efficiently improved. Another reason is that MLP kernel have higher accuracy compared to other kernels.

In order to avoid bias created in case participation between the training and testing sub-datasets from a limited small dataset, we applied a standardized leave-one-case-out (LOCO) cross-validation method to train and test this SVM model to classify between the malignant and benign pancreatic tumors. Using the LOCO method, radiomics feature vectors computed from 76 tumors are used to train SVM model and the radiomics feature vector of one remaining tumor that is not involved in training process is tested by the trained model. Such LOCO process is iteratively performed 77 times. Thus, each of 77 pancreatic tumors are independently tested once by the SVM models trained using other 76 pancreatic tumors.

3 Results

First, in evaluating the performance of pancreatic segmentation, several popular evaluation indices including Dice similarity coefficient (DSC) and pixel-wise accuracy were computed and analyzed. DSC can well reflect the overlap ratio between two sample sets under the comparison. In detail, the DSC of the pancreas segmentation model is defined as follows: $DSC = \frac{2 (| A ⋂ B |)}{| A | + | B |}$ where A and B represent the outputs of pancreas segmentation network and manual segmentation of the pancreas region depicting on CT image, respectively. Pixel-wise accuracy measures and reports the percentage of pixels in the ROI which are correctly classified in the image. The data analysis results show that our CNN-based image segmentation model yields pancreas region segmentation accuracy of DSC = 76.54±9.16 and pixel-wise accuracy of 73.65±2.31 when applying to the pancreas CT image dataset downloaded from TCIA [15].

Next, the SVM-generated classification scores of all 77 study cases are analyzed using a standard ROC-based statistical data analysis method (using ROCKIT program, https://metz-roc-uchicago.edu/MetzROC/software) to assess diagnostic or tumor classification performance. The SVM-generated tumor classification scores range from 0 to 1 in which the higher score indicates the higher likelihood of the tumor being malignant, while the lower score tends to be benign. The area under ROC curve (AUC) is used as an index to evaluate the overall performance of the SVM model to classify between the malignant and benign pancreatic tumors. Figure 6(a) shows the computed ROC curve of the SVM classification result applying to this dataset of 77 study cases. The computed corresponding area under ROC curve, AUC = 0.750 with 95% confidence interval (CI) range of [0.624, 0.885].

Fig. 6

(a) A ROC curve of SVM model to classify between malignant and benign pancreatic tumors, and (b) a confusion matrix of SVM classification result after applying an operation threshold of T = 0.5.

In addition, an operation threshold of tumor classification (T = 0.5) is applied to divide all testing cases into two classes of the malignant cases (T > 0.5) and benign cases (T≤0.5). After applying this operation threshold, a confusion matrix is generated (as shown in Figure 6(b)) and several additional classification performance indices are computed from the confusion matrix, which include overall classification accuracy, sensitivity and specificity. In detail, by defining malignant tumors as “positive” and benign tumors as “negative,” the classification accuracy of the SVM model performance is computed as follows: $Accuracy = \frac{TP + TN}{TP + TN + FP + FN} \times 100 %$ where TP is the classified true positive number (malignant tumors), TN is the true negative number (benign tumors), FP is the false positive number and FN is the false negative number, respectively. Meanwhile, the pancreatic tumor classification sensitivity and specificity are defined and computed as follows: $Sensitivity = \frac{TP}{TP + FN} \times 100 %$ $Specificity = \frac{TN}{TN + FP} \times 100 %$

Thus, based on the generated confusion matrix and computation results, SVM model classifies 33 tumors as the malignant tumors, which include 20 TP tumors and 13 FP tumors, and 44 tumors as the benign tumors, which include 36 TN tumors and 8 FN tumors, respectively. As a result, the overall accuracy of tumor classification is 72.7% (56/77). The tumor classification sensitivity and specificity are 60.6% (20/33) and 81.8% (36/44), respectively.

4 Discussion

In medical image informatics field, great research effort has been made in developing and evaluating CAD schemes of medical images aiming to assist clinicians (i.e., radiologists) more accurately and/or more efficiently detecting or diagnosing diseases and predicting disease prognosis or patients’ response to disease treatment [23, 24]. Recently, an emerging and promising concept of radiomics [25] has demonstrated its advantages to help expand CAD application fields and/or improve CAD performance [26]. In this study, we use radiomics concept and expand CAD approach to a new application field of classifying between malignant and benign pancreatic tumors detected from the CT images. This study has following unique characteristics and/or new observations to develop CAD schemes of medical images or help facilitate the progress of medical image informatics research.

First, since the pancreatic tumors are often subtle with low conspicuity and difficult to detect or diagnose, many previous CAD studies published in the literature (i.e., [11–13 , 23]) manually segmented pancreatic tumor regions and then computed image features from the segmented regions, which is not only time-consuming or tedious, and also suffers the potentially large inter-reader variability to manually define and draw boundary contour of the tumors with subtle boundary margins. This study demonstrates the feasibility of automatically segmenting pancreatic tumor regions from the CT images with higher accuracy by using a two-step approach to segment pancreas and tumor. A modified CNN-based image segmentation method is also applied and tested to segment pancreas regions in this study. By comparing to another state-of-the-art deep learning model [27] used to segment pancreas regions using the same dataset downloaded from TCIA, which reported DSC = 81.27±6.27, our simple CNN-model yields slightly lower or comparable segmentation accuracy. The pancreas region segmentation results can provide a good frame or boundary condition to help more efficiently and accurately segment pancreatic tumors, which aims to effectively reduce the risk of leakage of tumor segmentation.

Second, several publicly available software packages have been developed and widely used by many researchers to extract and compute variety of radiomics image features. Previous studies have also demonstrated the feasibility of selecting and identifying some effective radiomics image features that highly associate with cancer phenotype information and prognosis [28, 29]. In this study, we also used a popular radiomics feature computation software package (MaZda [18]) to extract and compute radiomics image features and then applied a unique three-stage sequential feature selection method to create an optimal feature vector including 12 radiomics features extracted from an initially large pool of 1,267 radiomics features. Our study demonstrates the feasibility of training and applying a simple and widely used machine learning model (SVM) using the identified optimal radiomics feature vector to classify between malignant and benign pancreatic tumors.

Third, the experimental results indicate that objective of this study has been achieved. We developed and tested a fully automated CAD scheme that enables to overcome the gap of previous CAD-related studies in detecting and diagnosing pancreatic cancer. The study demonstrates the feasibility to combine the automated pancreatic tumor segmentation algorithm and radiomics feature analysis-based machine learning approach into one integrated CAD scheme. Applying this new method to develop an integrated CAD scheme has significant advantages to help improve the efficacy of applying CAD schemes to assist radiologists detecting and diagnosing pancreatic cancer in future clinical practice.

Despite of the promising experimental results, we recognize that this preliminary study has several limitations, which need to be addressed and overcome in future studies. First, the image dataset used in this study is small, which cannot adequately represent the diverse population of pancreatic cancer in the clinical practice. In addition, the benign tumors included in our dataset are all high-risk tumors, which increases the difficult or subtle level of the dataset used in this study and thus produces the relatively lower AUC value (or classification accuracy) as comparing to some of previous studies reported in the literature (i.e., [11–13 , 23]). Second, due to the small dataset, the deep learning model for pancreas segmentation, radiomics feature selection method and SVM classifier training process may not be optimal. The scientific rigor or robustness of the CAD scheme (including the involved image processing algorithms, CNN-based image segmentation model and SVM classification model) has not been tested in this study. More studies are thus needed in the future to investigate how to optimize deep learning models using the relatively small medical image datasets and also achieve higher robustness of the models [30]. However, we believe that this is a valid preliminary study, which demonstrates the feasibility of our proposed CAD scheme that integrates the automated tumor segmentation algorithms and multi-feature fusion-based machine learning classifier together. The study also builds a promising scientific foundation for us to further develop and optimize the proposed CAD approach and schemes in future studies after collecting more diverse study cases and CT images from clinical practice.

Footnotes

Acknowledgments

This study is supported in part by research grant P30CA225520 from the National Cancer Institute to Stephenson Cancer Center, the University of Oklahoma Health Science Center.

References

Siegel

, Miller

K.D.

, Fuchs

and Jemal

, Cancer statistics, CA Cancer J Clin 71(1) (2021), 7–33.

Tummala

, Junaidi

and Agarwal

, Imaging of pancreatic cancer: An overview, J Gastrointest Oncol. 2(3) (2011), 168–174.

Costache

M.I.

, Costache

C.A.

, Dumitrescu

C.I.

, et al., Which is the best imaging method in pancreatic adenocarcinoma diagnosis and staging –CT, MRI or EUS? Curr Health Sci J, 43(2) (2017), 132–136.

Linguraru

M.G.

, Sandberg

J.K.

, Li

, et al. Atlas-based automated segmentation of spleen and liver using adaptive enhancement estimation, International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), (2009), 1001–1008.

Kainmüller

, Lange

and Lamecker

, Shape constrained automatic segmentation of the liver based on a heuristic intensity model, Proc. MICCAI Workshop 3D Segmentation in the Clinic: A Grand Challenge, (2007), 109–116.

Lee

, Kim

, Lee

, et al., Efficient liver segmentation using a level-set method with optimal detection of the initial liver boundary from level-set speed images, Comput Methods Programs Biomed 88(1) (2007), 26–38.

Cuingnet

, Prevost

, Lesage

, et al., Automatic detection and segmentation of kidneys in 3D CT images using random forests, International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), (2012), 66–74.

Shi

, Jiang

and Zheng

, A stacked generalized U-shape network based on zoom strategy and its application in biomedical image segmentation, Comput Methods Programs Biomed 197 (2020), 105678.

Tan

, Zhou

, Li

, et al., Automated vessel segmentation in lung CT and CTA images via deep neural networks, J Xray Sci Technol 29(6) (2021), 1123–1137.

10.

Saraswathi

H.S.

, Rafi

, Manjunath

K.G.

, et al., Review on computer aided diagnosis of pancreatic cancer using artificial intelligence system, In IEEE 2020 10th International Conference on Cloud Computing, Data Science & Engineering, (2020), 19565633.

11.

, Lin

, Hui

, et al., Computer-aided diagnosis for distinguishing pancreatic mucinous cystic neoplasms from serous oligocystic adenomas in spectral CT images, Technol Cancer Res Treat 15(1) (2016), 44–54.

12.

, Nie

, Reheman

, et al., Computer-aided diagnosis and staging of pancreatic cancer based on CT images, IEEE Access 8 (2020), 141705–141718.

13.

, Gu

, Hu

, et al., Applying a radiomics-based strategy to preoperatively predict lymph node metastasis in the resectable pancreatic ductal adenocarcinoma, J Xray Sci Technol 28(6) (2020), 1113–1121.

14.

Schwartz

L.H.

, Litiere

, de Vries

, et al., RECIST 1.1 –Update and clarification: from the RECIST committee, Eur J Cancer 62 (2016), 132–137.

15.

Roth

H.R.

, Farag

, Turkbey

E.B.

, et al., Data from pancreas-CT, The Cancer Imaging Archive, (2020). https://wiki.cancerimagingarchive.net/display/Public/Pancreas-CT

16.

Roth

H.R.

, Lu

, Farag

, et al., DeepOrgan: Multi-level deep convolutional networks for automated pancreas segmentation, International Conference on Medical Image Computing and Computer-Assisted Intervention –MICCAI, (2015), 556–564.

17.

Grady

, Random walks for image segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence 28(11) (2006), 1768–1783.

18.

Strzelecki

, Szczypinski

, Materka

and Klepaczko

, A software tool for automatic classification and segmentation of 2D/3D medical images, Nuclear Instruments & Methods in Physics Research A: Accelerators, Spectrometers, Detectors and Associated Equipment 702 (2013), 137–140.

19.

Wehrl

, General properties of entropy, Reviews of Modern Physics 50(2) (1978), 221.

20.

Fawcett

, An introduction to ROC analysis, Pattern Recognition Letters 27(8) (2006), 861–874.

21.

Kim

T.K.

, T test as a parametric statistic, Korean J Anesthesiol 68(6) (2015), 540–546.

22.

Osborne

M.R.

, Presnell

and Turlach.

B.A.

, On the LASSO and its dual, Journal of Computational and Graphical Statistics 9(2) (2000), 319–337.

23.

, Jiang

, Wang

, et al., An effective computer aided diagnosis model for pancreas cancer on PET/CT images, Comput Methods Programs Biomed 165 (2018), 205–214.

24.

Zargari

, Du

, Heidari

, et al., Prediction of chemotherapy response in ovarian cancer patients using a new clustered quantitative image marker, Phys Med Biol 63(15) (2018), 155020.

25.

Lambin

, Rios-Velazquez

, Leijenaar

, et al., Radiomics: Extracting more information from medical images using advanced feature analysis, Eur J Cancer 48 (2012), 441–446.

26.

Mirniaharikandehei

, Heidari

, Danala

, et al., Applying a random projection algorithm to optimize machine learning model for predicting peritoneal metastasis in gastric cancer patients using CT images, Comput Methods Programs Biomed 200 (2021), 105937.

27.

Roth

H.R.

, Lu

, Lay

, et al., Spatial aggregation of holistically-nested convolutional neural networks for automated pancreas localization and segmentation, Medical Image Analysis 45 (2018), 94–107.

28.

Khalvati

, Zhang

, Baig

, et al., Prognostic value of CT radiomic features in resectable pancreatic ductal adenocarcinoma, Sci Rep 9(1) (2019), 5449.

29.

Chen

, Liu

, Thai

T.C.

, et al., Developing a new radiomics-based CT image marker to detect lymph node metastasis among cervical cancer patients, Comput Methods Programs Biomed 197 (2020), 105759.

30.

Zuo

, Zheng

, He

, et al., Automated classification of papillary renal cell carcinoma and chromophobe renal cell carcinoma based on a small computed tomography imaging dataset using deep learning, Front Oncol 11 (2021), 746750.