CT radiomics to predict high‐risk intraductal papillary mucinous neoplasms of the pancreas

Abstract

Purpose

Intraductal papillary mucinous neoplasms (IPMNs) are radiographically visible precursor lesions of pancreatic cancer. Despite standard criteria for assessing risk, only 18% of cysts are malignant at resection. Thus, a large number of patients undergo unnecessary invasive surgery for benign disease. The ability to identify IPMNs with low or high risk of transforming into invasive cancer would optimize patient selection and improve surgical decision‐making. The purpose of this study was to investigate quantitative CT imaging features as markers for objective assessment of IPMN risk.

Methods

This retrospective study analyzed pancreatic cyst and parenchyma regions extracted from CT scans in 103 patients to predict IPMN risk. Patients who underwent resection between 2005 and 2015 with pathologically proven branch duct (BD)‐IPMN and a preoperative CT scan were included in the study. Expert pathologists categorized IPMNs as low or high risk following resection as part of routine clinical care. We extracted new radiographically inspired features as well as standard texture features and designed prediction models for the categorization of high‐ and low‐risk IPMNs. Five clinical variables were also combined with imaging features to design prediction models.

Results

Using images from 103 patients and tenfold cross‐validation technique, the novel radiographically inspired imaging features achieved an area under the receiver operating characteristic curve (AUC) of 0.77, demonstrating their predictive power. The combination of these features with clinical variables obtained the best performance (AUC = 0.81).

Conclusion

The present study demonstrates that features extracted from pretreatment CT images can predict the risk of IPMN. Development of a preoperative model to discriminate between low‐risk and high‐risk IPMN will improve surgical decision‐making.

1. Introduction and purpose

Pancreatic cancer is one of the most lethal cancers with an estimated 55,440 new cases and 44,330 deaths expected in 2018 in the United States.¹ Despite improvements in the understanding of the disease, the 5‐year survival rate is 7%¹ due to the late stage of diagnosis and ineffectiveness of current treatments. Early detection at a precancerous stage is widely viewed as the best tool to prevent pancreatic cancer.²

Intraductal papillary mucinous neoplasms (IPMNs) represent approximately 25% of all cystic neoplasms³ and are the only radiographically identifiable precursor of pancreatic cancer. IPMNs arise from ductal epithelium of main duct or branch duct.⁴ Approximately one‐third of resected IPMNs are associated with invasive carcinoma.⁵ The slow progression of this mucin‐producing neoplasm from low‐ to high‐grade dysplasia to invasive carcinoma offers an opportunity for early diagnosis and curative surgical treatment. Current laboratory, endoscopic, cytologic, and imaging technologies are limited in their ability to distinguish between IPMNs that are at low or high risk of harboring cancer. Prognostic markers of risk would select patients with high‐risk IPMNs for surgical resection, whereas patients with low‐risk IPMNs could be spared an invasive, highly morbid procedure and instead followed at regular intervals with imaging.⁶ Therefore, risk assessment is critical in the management of IPMNs.⁴

With the widespread use of abdominal cross‐sectional imaging, incidental pancreatic cysts are identified in up to 2.6% of all computed tomography (CT) scans.⁷ International consensus guidelines for the preoperative management of these patients based on clinical and radiographic criteria were published in 2006 and updated in 2012.^{8, 9} Larger size of the pancreatic duct or cyst, abrupt changes in duct caliber, and presence of solid‐enhancing component in the cyst are considered as the predictors of underlying malignancy.^{8, 9} However, the pleomorphic appearance of cystic lesions, ranging from simple well‐circumscribed cysts to infiltrative masses and every variation in‐between, presents a formidable challenge to reproducible visual assessments by radiologists. For example, radiographic evaluation has shown inconsistency in the presence of mural nodule or solid component.¹⁰ Still, visual assessment of IPMNs guides the decision for surgical resection. In the absence of main duct dilation (branch duct [BD]‐IPMN), management recommendations are more controversial.⁴ The yield from resecting BD‐IPMN is low; approximately 18% of resected tumors show malignancy on pathology report.⁴ Thus, a significant proportion of these patients undergo unnecessary and invasive surgery for benign lesions. Hence, preoperative markers of risk in BD‐IPMN are urgently needed.

Over the last decade, quantitative image analysis with a large number of features in combination with machine learning, popularly known as radiomics,¹¹ has shown the potential in oncologic imaging for prognosis, detection, and diagnosis of different cancers, including lung,¹² breast,¹³ prostate,^{14, 15} and pancreas.¹⁶ Fundamentally, radiomics is motivated by the observation that medical image features, such as intensity, shape, size, and/or texture, reflect underlying pathophysiology.¹⁵ Due to their successful use in many clinical applications, efforts have been made to standardize these features.¹⁷ Texture analysis provides a measure of intratumoral heterogeneity which can help in assessing tumor aggressiveness.¹⁵ However, radiomic studies for pancreatic cancer, in particular for IPMN risk assessment, are limited. Two recent preliminary studies demonstrated the potential for texture analysis to determine IPMN malignancy,^{18, 19} but the sample sizes were small with a low number of malignant BD‐IPMN. The authors acknowledged that the proposed prediction models likely suffered from overfitting. Furthermore, both papers analyzed mixed cohorts of branch duct (BD) and main duct (MD) IPMN; however, assessment of BD‐IPMN is more clinically relevant than MD‐IPMN since all MD‐IPMNs are recommended for resection.

The purpose of the present study was to define quantitative imaging markers for the objective assessment of BD‐IPMN risk. We based our image analysis methods on three radiographic observations related to increased cancer risk:

1. Enhancing solid component within the cyst and thickened/enhancing cyst walls are known visually assessed radiographic criteria of risk.⁴

2. Pancreas cancer is a whole‐gland process in which excess mucin production dilates the pancreatic duct; thus, hypoattenuated or dark regions within the pancreas parenchyma may correlate with risk.⁴

3. Fatty pancreatic infiltration is a risk factor for pancreatic precancerous lesions (PanIN);

thus, areas of fat or low attenuation within the pancreas may influence risk.²⁰

Therefore, we analyzed pancreatic cysts and parenchyma with new clinically and radiographically inspired features (RiFs) as well as texture features to predict cancer risk in BD‐IPMN.

Preliminary analysis of the pancreatic cyst region was presented at SPIE Medical Imaging in 2017.²¹ Building upon this preliminary work, the present paper proposes new features to quantify fatty infiltration of the pancreas, incorporates analysis of different regions of interest, and provides detailed analysis of imaging features in combination with clinical variables to demonstrate the improvement in performance of imaging predictors over clinical variables in predicting the risk of IPMN.

2. Materials and methods

An overview of the study design is shown in Fig. 1. Briefly, we explored and evaluated both novel radiographically inspired and standard texture features for the prediction of risk of IPMN. The prediction models were based on random forests (RF) and support vector machine (SVM) and were evaluated with tenfold cross‐validation. The features were also combined with clinical variables with known associations to risk.^{22, 23}

Figure 1.

Schematic for building prediction model of IPMN risk. [Color figure can be viewed at wileyonlinelibrary.com]

2.A. Study design and patients

A waiver of Health Insurance Portability and Accountability Act authorization and informed consent was obtained through the Institutional Review Board at our institution to conduct this study. Patients with portal venous phase CT scans and pathologically proven BD‐IPMN were from a consecutive series of patients with resected IPMN between 2005 and 2015 from a prospectively maintained database. A total of 103 patients met the inclusion criteria. Grade of dysplasia found at resection was assessed by an expert pathologist as part of routine clinical care and categorized into low risk and high risk based on standard criteria.²² Among all patients, 27 were with high‐risk IPMN.

Clinical variables were obtained from the database and by review of the electronic medical record. Five clinical variables: age at resection, cyst size, presence of solid component, presence of symptoms (abdominal pain or gastrointestinal disturbance), and gender — previously shown to be associated with IPMN risk²² — were included in our prediction model. In general, older symptomatic men with larger cysts and presence of solid component are at higher risk for invasive cancer. Among these, age and largest cyst size are continuous variables, and the presence of solid component, gender, and presence of any symptom are binary.

2.B. CT acquisition

Patients underwent contrast‐enhanced CT imaging as part of routine clinical management. The pretreatment postcontrast CT images were acquired following the administration of 150 mL iodinated contrast (Omnipaque 300, GE Healthcare, Princeton, NJ, USA) at 4.0 mL/s, on multidetector CT (Lightspeed 16 and VCT, GE Healthcare, WI). The scan parameters were as follows: pitch/table speed = 0.984–1.375/39.37–27.50 mm; autoMA 220–380; noise index 12.5–14; rotation time 0.7–0.8 ms; scan delay 80–85 s. Image analysis was performed on axial slices which are reconstructed at 2.5 mm intervals and 120 kVp. Pixel size was [0.8243, 0.8243] mm.

2.C. Image segmentation

After acquiring the images, the cyst and pancreas regions were manually segmented by an experienced radiologist, blinded to outcome, using Scout Liver (Analogic Corporation, Peabody, MA).

2.D. Extraction of novel radiographically inspired features (RiFs)

We designed four novel features based on radiographic observations that relate to high‐risk disease. We applied adaptive thresholding to capture high‐intensity pixels, representing solid component in cyst, and low‐intensity pixels, representing hypoattenuated dilation in parenchyma regions. The cyst or pancreas regions were initially smoothed to remove small fluctuations using an averaging filter of size 3 × 3 pixels. The filtered image was then thresholded to find the high/low‐intensity pixels of cyst/pancreas using a threshold. To compute the threshold, all the pixels within the region under consideration (pancreas/cyst) were sorted in ascending order of intensity, and the threshold was selected such that 90% of the total pixels were less than that intensity. The threshold is computed as $T{h}_i=\{{\mathbf{Is}}_i(n):n=\lceil 0.9*N_i\rceil ;{\mathbf{Is}}_i\text{contains ascending sorted intensity of}{\mathbf{I}}_i\}$ :

$$\begin{array}{cc}\hfill {\mathbf{IT}}_c& ={\mathbf{I}}_c\ge T{h}_c,\hfill \\ \hfill {\mathbf{IT}}_p& ={\mathbf{I}}_p\le T{h}_p.\hfill \end{array}$$ (1)

where i ∈ {c, p}, c and p represent cyst and pancreas, respectively. N _i is the total number of pixels in the segmented region I _i and IT _i is the thresholded image. I _p and I _c represent segmented cyst and pancreas regions, respectively. Pixels greater/less than the threshold are termed as enhanced pixels for cyst/pancreas region, respectively. We defined boundary and nonboundary regions of cyst/pancreas for extracting the RiFs based on radiologist's observations regarding the appearance of high‐risk IPMN as described in the introduction.^{4, 20} The boundary width was empirically selected as three pixels to quantify the enhanced wall that appears as thin ring‐like structure. Let N _i,b and N _i,in be the number of boundary pixels and nonboundary pixels of I _i , where boundary and nonboundary regions are mutually exclusive.

We extracted the following four RiFs from cyst and pancreas region.

2.D.1. Enhanced boundary fraction (EBF)

According to our observations 1 and 3, to quantify the enhancement of cyst wall or infiltration of pancreas boundary, EBF is defined as the ratio of enhanced pixels to the total number of pixels at boundary:

$${\text{EBF}}_i=\frac{E_{i,b}}{N_{i,b}},$$ (2)

where E _i,b represents the number of enhanced boundary pixels (i.e., boundary pixels having values greater/less than the threshold for cyst/pancreas).

2.D.2. Enhanced inside fraction (EIF)

EIF is the fraction of enhanced nonboundary pixels of IT _i to the total number of nonboundary pixels:

$${\text{EIF}}_i=\frac{E_{i,in}}{N_{i,in}}$$ (3)

where E _i,in represents the number of enhanced nonboundary pixels. The feature quantifies the high/low‐intensity pixels within cyst/pancreas, caused by solid‐enhancing component/mucin dilation, as we observed in 1 and 2.

2.D.3. Filled largest connected component fraction (FLCCF)

The largest enhanced area obtained by filling the largest enhanced component within the nonboundary region is termed as the filled largest connected component (FLCC). Let IT i contains K number of connected components cc ₁, cc ₂, ···, cc _k ; after filling the holes inside; then, cc _max is the largest region representing FLCC.

FLCCF is the ratio of area of FLCC to the area of nonboundary region:

$${\text{FLCCF}}_i=\frac{\text{Area}(c{c}_{\mathit{max}})}{N_{i,in}}.$$ (4)

2.D.4. Average‐weighted eccentricity (AWE)

Often one or more connected components may be found with size close to the largest connected component (FLCC), and consideration of only the largest component may miss some information. AWE is based on the area of each connected component defined as:

$${\text{AWE}}_i=\frac{\sum _{k=1}^K\text{Area}(c{c}_k)*\text{Ecc}(c{c}_k)}{\sum _{k=1}^K\text{Area}(c{c}_k)},$$ (5)

where Ecc(cc _k ) represents the eccentricity of cc _k . We extracted the RiFs from each image slice and averaged to obtain the final RiFs for the image. An example of high‐and low‐risk lesions with their enhanced component at boundary and nonboundary regions for both cyst and pancreas region are shown in Figs. 2 and 3, respectively. These figures demonstrate areas of enhanced pixels in the cyst and pancreas for the patient with high‐risk IPMN compared to that of the low‐risk IPMN.

Figure 2.

An example of the enhanced component at boundary and nonboundary regions of the cyst for low‐risk (top panels) and high‐risk (bottom panels).

Figure 3.

An example of the enhanced component at boundary and nonboundary regions of the pancreas for low‐risk (top panels) and high‐risk (bottom panels).

2.E. Extraction of texture features

In this study, we extracted several intensity and orientation‐based texture features. Based on the assumption that image texture has two components, spatial structure (pattern) and strength (contrast), illumination invariant local binary pattern (LBP) were introduced. LBP quantifies the structural information by assigning a value for each pixel by thresholding the 3 × 3 neighbors with the center pixel and computing the decimal value corresponding to the generated eight bit stream. We computed histograms of uniform LBP (ULBP) and rotation invariant ULBP (RI‐ULBP)^{24, 25, 26}, two modified operators, which remove less frequently occurring nonuniform patterns and provide rotation invariant patterns, respectively. Four statistical measures (standard deviation, skewness, kurtosis, and entropy) of ULBP and RI‐ULBP were used as features along with the Fourier descriptors extracted from the RI‐LBP histogram.²⁷ A set of 115 LBP features were thus constructed: 59 ULBP (L ₁–L ₅₉), 10 RI‐ULBP (L ₆₀–L ₆₉), 8 statistical (L ₇₀–L ₇₇), and 38 frequency (L ₇₈–L ₁₁₅) descriptors.

Orientation of tissue patterns may be associated with cancer. Hence, to characterize oriented patterns present in an image, two angle co‐occurrence matrices (ACMs) were designed to capture the directional edge patterns of a region via computing the probability of occurrences of two orientation angles at a particular direction and distance.^{28, 29} The first matrix (ACM1) was designed using only gradient angle, whereas ACM2 was formed using both gradient angle and magnitude. Since ACM2 carries both angle and magnitude information, we used ACM2 to analyze the oriented patterns. The gradient information was extracted with Sobel operator of size 3 × 3 pixels. To achieve rotation invariant features, ACMs were computed in four directions (0°, 45°, 90°, and 135°) and formed a single matrix with averaging. Sixteen rotation invariant features including 14 Haralick statistics, cluster shade, and cluster prominence (A ₁–A ₁₆) were extracted from the resultant matrix^{30, 31, 32}.

The features were extracted from each axial slice and averaged over all the slices to obtain one single value for each feature. A total of 131 texture features (Tx) were derived from each cyst and pancreas regions. Texture differences in the cyst region for two exemplar high‐ and low‐risk IPMNs are shown in Fig. 4. The feature matrices show wider spread of intensity values and oriented patterns, which suggest more heterogeneity in high‐risk IPMN.

Figure 4.

Tumors with rendered texture features for patients low‐risk (top row) and a high‐risk (bottom row) IPMN. For better visualization of the features, the angle co‐occurrance matrix (ACM) is represented with pseudo image. [Color figure can be viewed at wileyonlinelibrary.com]

2.F. Feature selection and classification

To evaluate the performance of the proposed features, the RF classifier and SVM with radial basis function were selected due to their efficient performance for high‐dimensional data.³³ The parameters of the classifiers were selected by optimizing the model with the training data. To account for imbalanced data, we used cost‐sensitive learning, a technique preferred over sampling methods for the classification of imbalanced data.³⁴ In this technique, instead of assigning equal misclassification costs for both classes, higher cost is assigned for the class with fewer data. In our case, increased cost is associated with false‐negative (low‐risk) cases than false‐positive cases (high‐risk). The cost was selected experimentally with the training data. A two‐stage feature selection technique was used to reduce dimensionality by selecting the most discriminatory features and removing redundant features with univariate analysis followed by applying forward selection‐based multivariate analysis combining features. Since Wilcoxon rank‐sum test (WRST)‐based feature selection yields better predictability compared to other techniques,³⁵ it was first applied to remove the features which do not show significant association with the risk (P > 0.05) and to rank the features based on discriminatory power. Highly correlated features with low discriminatory power were then removed, based on the empirical evidence of their adverse effect on RF classifiers.³³ The threshold for removing correlated features was selected experimentally using the training data only. The threshold value was varied from 0.7 to 1.0 and the value which provided best area under receiver operating characteristic (ROC) curve (AUC) for the training data was selected as the optimum threshold. The set of first S‐ordered features, which provides best predictive performance in terms of AUC obtained for the training data, was then selected as the final features set. An ensemble method of feature selection was also employed to investigate the robustness of the features in selection where five individual feature selection techniques, namely P‐value from WRST, Pearson correlation, and Spearman correlation, β coefficient of logistic regression, and variable importance obtained with RF classifier were used. The features were ranked based on these techniques and ensembles to obtain the final ranking of the features. The first S number of ranked features that obtained best AUC on the training data was then selected as the final features set of the ensemble method.

2.G. Evaluation

The proposed method was implemented in MATLAB, version R2015a (Natick, MA, USA). The association of individual features with risk was estimated using P‐value computed with WRST. The performance of models designed with different combinations of features was evaluated with a tenfold cross‐validation technique. For small datasets, hold‐out validation is not feasible, and in such circumstances, cross‐validation is an effective choice.³⁶ In tenfold cross‐validation, observations are randomly divided into ten groups. Each of the groups is used exactly once as the test set while the others are used for training (i.e., to select the important features and to design the predictive model with the selected features). In the present study, the method was repeated ten times to avoid any bias caused by random splitting of data. The performance of models was evaluated with the commonly used metrics AUC, sensitivity (Sn), specificity (Sp), and positive and negative predictive value (PPV and NPV). However, values of Sn, Sp, PPV, and NPV vary depending on the selection of operating point. In this study, we computed values, placing equal importance on both classes to maximize Sn and Sp.

3. Results

We defined four RiFs (EBF, EIF, FLCCF, and AWE) and texture features that were extracted from 103 CT images from patients with confirmed IPMN. The performance of the individual RiFs was evaluated using AUC and P‐value obtained with WRST. Performance of imaging features extracted from the cyst and pancreas regions were investigated individually as well as in combination to explore the effect of different regions on prediction accuracy. Finally, imaging features were combined with clinical variables to improve predictive power.

3.A. Performance of RiFs

Performance of individual RiFs for predicting risk is summarized in Table 1. Notably, a significant association was observed between high‐risk and enhancing components present at the cyst boundary (EBF _c ) as well as at the inner pancreas (EIF _p ). Furthermore, areas of enhanced pixels in the cyst (FLCCF _c ) and pancreas (FLCCF _p ) were larger in high‐risk patients. Multiple areas of increased brightness in the cyst (AWE _c ) and dark regions in the pancreas (AWE _p ) showed association with high risk. These are consistent with the radiology literature.⁴ Although the combination of all RiFs from each region (cyst/pancreas) provides better performance than their individual features with RF, for SVM performance does not improve. The prediction model with all RiFs delivered significant performance improvement over individual groups, achieving an AUC of 0.77 ± 0.037 with ensemble feature selection and RF classifier (Table IV).

Table 1.

Performance of individual RiFs

Feature	Cyst region		Pancreas region
	P‐value	AUC	P‐value	AUC
EBF	0.001	0.703	0.218	0.580
EIF	0.451	0.549	0.014	0.659
FLCCF	0.001	0.713	0.049	0.628
AWE	0.024	0.647	0.002	0.704

3.B. Performance of texture features

For individual texture features extracted from the cyst, 12 features showed statistically significant association (P < 0.05) with risk. On the contrary, only one texture feature from pancreas was significant. Consequently, we excluded texture features of pancreas from further analysis. The performance of Tx from cyst regions was evaluated after designing the prediction model using discriminating features, selected with the feature selection method. The results in Tables 2, 3, 4, 5 demonstrate better performance of Tx compared to RiFs from the cyst region using SVM and vice versa using RF. An AUC of 0.74 is observed with SVM and ensemble feature selection technique.

Table 2.

Performance of features, selected with P‐value and forward selection, obtained using tenfold cross‐validation and random forest classifier. “*” represents no feature selection was applied

Feature set	AUC	Sn	Sp	PPV	NPV
RiFs from cyst	0.71 ± 0.011	0.72 ± 0.088	0.68 ± 0.103	0.46 ± 0.054	0.88 ± 0.019
RiFs from pancreas	0.73 ± 0.011	0.63 ± 0.048	0.80 ± 0.044	0.54 ± 0.059	0.86 ± 0.014
All RiFs	0.76 ± 0.012	0.72 ± 0.089	0.73 ± 0.081	0.49 ± 0.047	0.88 ± 0.022
Tx from cyst	0.70 ± 0.021	0.56 ± 0.092	0.81 ± 0.101	0.54 ± 0.091	0.84 ± 0.013
All imaging (Tx from cyst and all RiFs)	0.77 ± 0.071	0.74 ± 0.086	0.76 ± 0.088	0.54 ± 0.061	0.89 ± 0.023
Clinical*	0.68 ± 0.012	0.62 ± 0.158	0.70 ± 0.142	0.45 ± 0.069	0.85 ± 0.033
Clinical, all RiFs	0.80 ± 0.025	0.89 ± 0.052	0.57 ± 0.046	0.42 ± 0.024	0.94 ± 0.026
Clinical, all imaging	0.78 ± 0.013	0.70 ± 0.066	0.81 ± 0.059	0.57 ± 0.062	0.88 ± 0.018

Table 3.

Performance of features, selected with P‐value and forward selection, obtained using tenfold cross‐validation and support vector machine. “*” represents no feature selection was applied

Feature set	AUC	Sn	Sp	PPV	NPV
RiFs from cyst	0.72 ± 0.012	0.76 ± 0.082	0.68 ± 0.059	0.46 ± 0.027	0.89 ± 0.028
RiFs from pancreas	0.69 ± 0.034	0.70 ± 0.058	0.64 ± 0.07	0.41 ± 0.04	0.85 ± 0.05
All RiFs	0.73 ± 0.010	0.56 ± 0.142	0.82 ± 0.14	0.52 ± 0.06	0.84 ± 0.03
Tx from cyst	0.73 ± 0.015	0.72 ± 0.108	0.72 ± 0.076	0.48 ± 0.044	0.88 ± 0.03
All imaging (Tx from cyst and all RiFs)	0.75 ± 0.008	0.76 ± 0.068	0.70 ± 0.066	0.48 ± 0.04	0.89 ± 0.02
Clinical*	0.65 ± 0.011	0.62 ± 0.019	0.74 ± 0.03	0.46 ± 0.03	0.85 ± 0.01
Clinical, all RiFs	0.76 ± 0.023	0.62 ± 0.021	0.86 ± 0.03	0.62 ± 0.05	0.86 ± 0.01
Clinical, all imaging	0.76 ± 0.025	0.80 ± 0.089	0.59 ± 0.08	0.45 ± 0.04	0.88 ± 0.02

Table 4.

Performance of features, selected with ensemble feature selection, obtained using tenfold cross‐validation and random forest

Feature set	AUC	Sn	Sp	PPV	NPV
RiFs from cyst	0.73 ± 0.013	0.61 ± 0.061	0.81 ± 0.068	0.55 ± 0.063	0.86 ± 0.013
RiFs from pancreas	0.74 ± 0.01	0.83 ± 0.16	0.58 ± 0.13	0.42 ± 0.05	0.92 ± 0.05
All RiFs	0.77 ± 0.037	0.83 ± 0.076	0.67 ± 0.098	0.48 ± 0.054	0.92 ± 0.025
Tx from cyst	0.70 ± 0.010	0.76 ± 0.061	0.64 ± 0.086	0.44 ± 0.041	0.89 ± 0.018
All imaging (Tx from cyst and all RiFs)	0.78 ± 0.006	0.68 ± 0.07	0.84 ± 0.087	0.62 ± 0.104	0.88 ± 0.015
Clinical, all RiFs	0.81 ± 0.014	0.84 ± 0.072	0.70 ± 0.061	0.50 ± 0.032	0.93 ± 0.026
Clinical, all imaging	0.78 ± 0.008	0.73 ± 0.043	0.79 ± 0.054	0.56 ± 0.059	0.89 ± 0.011

Table 5.

Performance of features, selected with ensemble feature selection, obtained using tenfold cross‐validation and support vector machine

Feature set	AU C	Sn	Sp	PPV	NPV
RiFs from cyst	0.71 ± 0.016	0.77 ± 0.103	0.64 ± 0.104	0.44 ± 0.049	0.89 ± 0.032
RiFs from pancreas	0.67 ± 0.009	0.72 ± 0.148	0.57 ± 0.161	0.40 ± 0.058	0.88 ± 0.040
All RiFs	0.72 ± 0.014	0.71 ± 0.08	0.67 ± 0.078	0.45 ± 0.038	0.87 ± 0.019
Tx from cyst	0.74 ± 0.010	0.78 ± 0.059	0.71 ± 0.059	0.49 ± 0.038	0.90 ± 0.017
All imaging (Tx from cyst and all RiFs)	0.78 ± 0.012	0.74 ± 0.065	0.79 ± 0.064	0.56 ± 0.062	0.90 ± 0.017
Clinical, all RiFs	0.74 ± 0.007	0.65 ± 0.033	0.80 ± 0.029	0.54 ± 0.025	0.86 ± 0.010
Clinical, all imaging	0.79 ± 0.006	0.78 ± 0.043	0.76 ± 0.046	0.54 ± 0.040	0.91 ± 0.014

3.C. Analysis of imaging features in combination

Combination of all imaging features (i.e., Tx from cyst, RiFs from cyst, and pancreas) obtained better performance than individual groups. The combination achieved an AUC of 0.78 ± 0.006 with Sn and Sp of 0.68 ± 0.070 and 0.84 ± 0.087, respectively, with ensemble feature selection followed by RF‐based classification (Table 4, all imaging).

3.D. Analysis of imaging features with clinical variables

The association of risk with demographic and clinical characteristics are described in Table 6. Although these variables are usually statistically significant,²² only the presence of symptoms and gender were correlated with risk in our dataset.

Table 6.

Patient demographics and clinical characteristics between high‐ and low‐risk groups

	Features all patients	High‐risk	Low‐risk	P‐value
Median age (range), yr	68.0 (44–90)	65.0 (46.0–82.0)	68.5 (44–90)	0.5812
Median largest cyst size (range), cm	2.9 (1–6)	3.2 (1.5–5.2)	2.8 (1–6)	0.2358
Presence of solid component	40 (38.8%)	14 (51.9%)	26 (34.2%)	0.1089
Presence of symptoms	44 (42.7%)	18 (66.7%)	26 (34.2)	0.0036
Male	47 (45.6%)	17 (63.0%)	30 (39.5%)	0.0366

Results of combining imaging features with clinical variables are summarized in Tables 2, 3, 4, 5. The combination of clinical features with all RiFs provided the best AUC of 0.81 ± 0.014 using ensemble feature selection and RF classifier among all different prediction models analyzed in this study (Table 4, Clinical All RiFs). Although the fusion of clinical variables with all imaging features (RiFs and Tx) led to poorer AUC, Sn, and NPV, it significantly increased Sp and PPV for RF classifier. On the contrary, Sn and NPV increased for SVM (Table 5). A model designed with the clinical variables alone clearly established the efficacy of imaging features over clinical features (see Tables 2, 3, 4, 5). The ROC curves corresponding to Tables 2, 3, 4, 5 are shown in Fig. 5.

Figure 5.

ROC curves obtained with imaging features (first column) and combination of imaging features with clinical variables (second column) corresponding to Tables 2, 3, 4, 5. [Color figure can be viewed at wileyonlinelibrary.com]

4. Discussion

This retrospective study investigated the predictive power of quantitative CT imaging features for the risk assessment of BD‐IPMN and demonstrated strong potential. Categorization of IPMNs into high and low risk, prior to treatment, is clinically important since their preoperative discrimination would allow patients with high‐risk IPMN to undergo resection prior to the development of an incurable disease and enable patients with low‐risk IPMN to avoid unnecessary surgery for benign disease. However, currently, no successful method exists to distinguish them.

Among all imaging features, the RiFs — EBF _c , FLCCF _c , and AWE _p — were often selected with different texture features using the feature selection method. It was observed that the proposed radiographically inspired features outperformed standard texture features. From experiments with all imaging features, RiFs were selected more often than texture features. Three of eight (37.5%) RiFs were selected at each iteration and fold (100%), whereas only 16 were selected from 131 texture features (12.2%) with an occurrence rate of 100% for three features (2%) and more than 95% for six features (5%). RiFs are likely more adaptive than texture features due to the adaptive thresholding utilized in their design. Furthermore, texture features are more sensitive to noise, and the addition of more texture features does not improve model performance. RiFs are based on established radiographic criteria, whereas the underlying biological meaning of texture features is not well elucidated.

In addition to the 131 texture features, included in this study, we investigated 14 Haralick features extracted from GLCM³¹ and five intensity histogram‐based features (mean, standard deviation, skewness, kurtosis, and entropy). However, none of these features showed significant performance in univariate analysis with WRST in our dataset. A list of the features with their P‐values are provided in Table S1. Moreover, cyst volume was computed and analyzed; however, no significant association with IPMN risk was found (P = 0.3412). Therefore, we excluded those features from the prediction model.

Based on the distribution of data, one classifier can work better than another. For the present study, we designed models using two well‐known classifiers — RF and SVM. It is observed that the models designed with RF provided better performance than SVM, except for texture features extracted from cyst region. Since the ensemble method selects features based on multiple selection algorithms, it was expected to show better outcome than other techniques. However, both the feature selection methods, used herein, demonstrate comparable performance with slightly better performance with ensemble feature selection. We also investigated elastic net for feature selection, which did not yield satisfactory performance. The results are provided in Tables S2 and S3.

The high NPV obtained with our methods is clinically significant; better detection of patients with low‐risk IPMNs (high NPV) would prevent unnecessary resection of patients unlikely to develop pancreas cancer. Short‐term morbidity after pancreas resection remains high at 30–40%, and therefore, preventing unnecessary surgery is critical.^{37, 38}

Quantitative CT analysis for IPMN risk assessment is limited. Hoffman et al.³⁹ extracted a few intensity histogram‐based statistical features from MRI of 18 BD‐IPMN patients and demonstrated that entropy could be prognostic of malignancy. In a concept‐of‐proof study with 38 patients, including 20 benign and 18 malignant IPMNs, Permuth et al.¹⁹ showed the potentiality of texture analysis to discriminate benign and malignant IPMNs with an AUC of 0.77. Hanania et al.¹⁸ have also drawn similar conclusion with 34 high‐risk and 19 low‐risk IPMN patients which are consistent with the present study. However, all these methods were investigated with small sample sizes. The authors too acknowledged that the proposed prediction models likely suffered from overfitting. Furthermore, the significant performance of Permuth and Hanania et al. may be due to high‐risk profile of main duct disease, since they used both BD‐ and MD‐IPMNs.⁹ Instead, the proposed study investigated only BD‐IPMN patients and achieved an AUC of 0.81 with a dataset of 103 patients. Assessment of BD‐IPMN is clinically more relevant than MD‐IPMN as resection is recommended for all MD‐IPMNs.⁴

To evaluate the robustness of the proposed method with respect to variation in segmentation, we varied the manually drawn contour by applying an active contour‐based segmentation method.⁴⁰ No significant variation in performance is observed when the segmentation methods were compared (see Table S4).

A limitation of our study is that manual segmentation of the pancreas and cyst regions requires a radiologist which prevents fully automatic implementation of the model in its current form. The present study depends on the pathological diagnosis of risk, and thus, our cohort is restricted to resected patients. In addition, the study is limited by a small dataset and lack of external data for validation. Our upcoming multi‐institutional prospective trial following patients with pancreatic IPMN will yield biopsy material and allow for external validation. In the future, we will expand our patient cohort, perform a detailed study to address the impact of segmentation variability (if any) with multiple radiologists on performance, and design a fully automatic prediction model via automatic segmentation of pancreas and cyst regions. Moreover, the present study performed a 2D image analysis for IPMN‐risk stratification. In the future, we will investigate 3D CT volume via defining the texture and RiFs in 3D.

5. Conclusions

We investigate quantitative CT analysis to predict patients with high‐risk IPMN in a series of 103 resected patients with pathologically proven BD‐IPMN and show their efficacy. We identify novel RiFs that demonstrate good predictive power. The combination of these features with standard texture features slightly improved performance. Among models designed with different feature sets and classifiers, the RF designed with the combination of new RiFs with clinical variables provides the best predictive performance. Our findings lay the groundwork for development of a preoperative model to discriminate between low‐risk and high‐risk IPMN. Work is in progress to evaluate our proposed features with data provided by outside institutions followed by independent validation.

Conflicts of interest

The authors have no relevant conflicts of interest to disclose.

Supporting information

Table S1: P‐values obtained with GLCM‐ and intensity histogram‐based features from cyst and pancreas region.

Table S2: Performance of features, selected with elastic net, obtained using tenfold cross‐validation and support vector machine.

Table S3: Performance of features, selected with elastic net, obtained using tenfold cross‐validation and random forest classifier.

Table S4: Performance comparison between imaging features extracted from two segmentation techniques using tenfold cross‐validation, ensemble feature selection, and random forest. The P‐value obtained between the two ROC curves indicates no significant difference between them (P > 0.05).