Evaluation of radiomics and radiologic signatures for diagnosis and clinical outcomes in pancreatic adenocarcinoma and pancreatic neuroendocrine: A pilot study

Abstract

Background

Biopsy is the only definitive method for diagnosing pancreatic ductal adenocarcinoma(PDAC) in the preoperative setting. With the increasing use of neoadjuvant therapy, alternative diagnostic modalities would aid in cases where fine-needle aspiration(FNA) is inconclusive or challenging. We hypothesized that radiomic signatures generated from computed tomography(CT) scans can help diagnose PDAC with high fidelity and radiologic/radiomics features have utility in predicting clinical outcomes.

Methods

Patients with PDAC who underwent oncologic pancreatectomy from 2017–2019 were selected. Preoperative CT scans were reviewed by two body radiologists and two pancreatic surgeons. Tumors were segmented using two radiomics software platforms(Platform A&B).

Results

Twenty-five patients were included (n = 20 PDAC,n = 5 pancreatic neuroendocrine tumors(pNET)). Neoadjuvant chemotherapy was administered in 80% of PDAC patients, and 60% underwent pancreaticoduodenectomy. Platform A extracted 455 independent radiomics features and Platform B extracted 2078 radiomics features corresponding to shape, texture, filter, and intensity. One principal component from the shape feature group showed significant discrimination between PDAC and pNET, within both Platform A and B(area under the curve(AUC) = 0.81,p = 0.038 and AUC=0.89,p = 0.017, respectively). Additionally, a gray-level co-occurrence matrix sum entropy feature identified with Platform B differentiated malignant pancreatic tissue from normal parenchyma and PDAC from normal gland(p < 0.05).

Conclusion

Radiologic/radiomic features are reliable tools for PDAC diagnosis and associate with clinical outcomes. An improved understanding of PDAC radiomics may have implications for prognosis, margin status, and treatment response assessments.

Precis: Radiologic/radiomic features are reliable tools for PDAC diagnosis and associate with clinical outcomes. An improved understanding of PDAC radiomics may have implications for prognosis, margin status, and treatment response assessments.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is the eleventh most common malignancy in the United States yet the third most lethal. The incidence of pancreatic cancer has increased in the United States, with an estimated 56,700 new cases and 45,700 deaths per year.^1,2 Resection to negative margins with the addition of modern multidisciplinary therapy regimens provide the best survival outcomes, with a median overall survival as high as 54.4 months in a recent trial. ³ Traditionally, cytotoxic chemotherapies have been administered following surgical resection to eradicate microscopic residual disease and decrease the likelihood of recurrence. Chemotherapy increasingly has been administered prior to surgery (neoadjuvant chemotherapy) in pancreatic cancer. While no level 1 data exists demonstrating the superiority of one treatment sequence over the other, supporters of the neoadjuvant approach point to the testing of disease biology, better compliance with systemic regimens, and improved margin negativity as potential advantages.

In most circumstances, administration of neoadjuvant therapy requires a histologic diagnosis of adenocarcinoma. This is accomplished by endoscopic ultrasound with fine-needle aspiration biopsy or image-guided percutaneous biopsy, though less frequently. These invasive methods, however, are operator dependent and not without risk. ⁴ Pancreatic fistula, pseudoaneurysm, pancreatitis, and the potential for cancer dissemination have been reported as morbidities of pancreatic biopsy. ⁵ In addition, certain tumors can be technically challenging to target for biopsy via these methods, and the limited availability of advanced endoscopists and/or experienced interventional radiologists remains a diagnostic challenge in many medical centers. ⁶ In addition the tissue amount is limited and rarely can be sent for next generation sequencing or genomic evaluation based on the cellularity. Less invasive, reliable alternatives and/or complimentary to tissue diagnosis would be of great value for the diagnosis and treatment of pancreatic cancer.

Radiomics is an emerging technology that utilizes advanced imaging features generated from imaging studies to describe tumors objectively and quantitatively. Radiomics features are numeric data points within imaging files that correspond to textural and shape characteristics within a cross-sectional image. Radiomics features have recently been demonstrated in several studies to correspond with tumor genotype and predict outcome in certain disease sites, including lung and liver cancers.^7,8 Radiomics techniques, however, have not been evaluated in pancreatic cancer with the same degree of validation. Application of these techniques within pancreatic cancer may improve diagnostics with a high level of fidelity without the requirement for an invasive tissue biopsy. Furthermore, radiomics features may have utility in the prediction of response to therapy and overall prognosis. More subjective metrics of cross-sectional imaging, such as delta and interface score, have already proved reliable in predicting outcomes. ⁹ However, such techniques are subject to the variability of interpretation and are limited to mainly dichotomous inputs. Radiomics analysis has the benefit of objectivity and employing a much greater breadth and variability of data for use in predictions.

Radiomics techniques have both diagnostic and prognostic applications in the management of several cancers. The purpose of this pilot study aims to evaluate radiomics techniques in pancreatic adenocarcinoma and compare it to existing subjective metrics to see if radiomic techniques can provide diagnostic comfimation without tissue biopsy. We hypothesize that radiomics features can be utilized to reliably diagnose pancreatic cancer from imaging and has potential utility in predicting outcomes in the disease.

Methods

This study protocol was reviewed and approved by University of Louisville Institutional Board, approval number 06.0326.

Inclusion criteria

Sequential patients diagnosed with pancreatic tumors who underwent oncologic pancreatectomy between January 2017 and December 2019 were eligible for inclusion in this study. This was a sub-study of a larger prospective data collection database that was initial IRB(June 2006) approved, and continued approval at the University of Louisville School of Medicine. Written informed consent was obtained from the parent/legal guardian of participants prior to the study. Patients included must have been staged with a three-phase pancreatic protocol (1 mm cut) computed tomography (CT) scan prior to their operation. Patients with poor quality imaging, severe scatter artifacts from foreign bodies, and cystic masses were excluded. Neoadjuvant therapy status was not considered in patient selection. The study was conducted with approval from the institutional review board and followed all informed consent practices and prospective data collection.

Data collection

Patient data was collected from our prospectively maintained institutional database. Variables collected included demographic information, CA 19–9 levels, tumor characteristics, intraoperative and pathologic variables. Long-term recurrence, adjuvant therapy, and mortality information were also collected. Follow-up was obtained from all patients and was updated until December 2020.

Images were analyzed by two separate radiomics platforms, Platform A (HealthMyne (Madison, WI)) and Platform B (Tempus, Chicago IL). Each platform included a combination of manual radiology review and AI-assisted extraction of radiomics features and are described in more detail in the following sections.

In addition, A third non-radiomics based technique was used to evaluate the tumor delta. CT scans were reviewed by two independent experienced pancreatic surgeons to designate tumor delta in the method described by Koay et al.. ⁹ For specifics at baseline a standard-of-care pancreas protocol CT, high-delta PDAC tumors exhibit an abrupt change—or delta—in Hounsfield units (HU) between the visualized tumor and normal pancreas, and low-delta PDAC tumors do not exhibit such a change. After preoperative therapy, tumors with a type I response remain or become well defined at the interface of tumor and parenchyma, whereas those with a type II response become less defined at the interface. Previously these CT-based biomarkers associate with pathologic features of PDAC, such as the extent of stromal reaction. Recent literature consistently supports a pilot sample sizes in the range of 20–30 participants as sufficient to: Assess feasibility metrics (radiomic protocol compliance), characterize image acquisition and contouring procedural variability, and provide estimate effect size and variance for future power calculations.^10,11

Radiomics platform A feature generation

Images were copied into Platform A's laptop computer for initial evaluation and segmentation. Segmentation of tumors and normal pancreatic parenchyma were performed independently by a blinded experienced pancreatic surgeon and body radiologist. Segmentations and radiomics were done with a HealthMyne (Madison, WI) (HM) software package. The entire lesion, including extrapancreatic extension, was annotated as part of the tumor. Vessels were subtracted from the final contours, as well as foreign bodies (i.e., stents). Radiomics data points were generated from this contouring. Contouring was mostly carried out on portal venous phase imaging.

Radiomics Platform A extracted 455 independent radiomics features that corresponded to shape (45), intensity (65), texture (288), and filter (57) features, with several non-applicable or non-diagnostic features. Texture features included gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), neighboring gray tone difference matrix (NGTDM), and neighboring gray-level dependence matrix (NGLDM) features. Filter features were derived using Wavelet and Log2D filters. This platform also extracted numerous lung features and Hounsfield unit (HU) threshold features not applicable to pancreatic analysis.

Radiomics platform B feature generation

Images were received by Radiomics Platform B via a secure transfer protocol as DICOM files and de-identified via Google Cloud's Healthcare API. These files were subsequently converted into a structured database for research purposes, using internal tools for data organization, which included a conversion to NIfTI (Neuroimaging Informatics Technology Initiative), a standard imaging research file format.

Axial orientation abdomen CT exams with portal venous and arterial contrast phase were annotated by two radiology fellows with six years of clinical experience. Each followed a standard protocol to mark the pancreas organ and pancreatic mass (Figure 1), generating regions of interest (ROI). Volume annotation was performed using the open-source software package ITKSnap. The annotation protocol emphasizes high attention to detail at contour boundaries and includes a second-phase review by a board-certified abdominal imaging radiologist to identify errors that required revision. Unclear pancreatic masses were more closely reviewed via a discussion and consensus with a pancreatic treating surgeon. Normal tissue contours were generated from the pancreas and tumor ROIs by removing any volume inside, or within 1 cm of the tumor from the full pancreas ROI.

Figure 1.

An example annotation of a single patient's pancreas and pancreatic tumor. Subfigures 1a - 1c show descending slices of the annotation, with the pancreas region of interest (ROI) highlighted in purple, the tumor ROI highlighted in light blue, and the aorta artery highlighted in green. Subfigure 1d shows a 3D surface map of the pancreas in gray, and the tumor in light blue.

Radiomics platform A and B analysis

Two primary methods were used on the radiomic features (described in Supplemental Table 1) from both Radiomics Platforms to search for imaging features which separate pancreatic adenocarcinoma and pancreatic neuroendocrine tumor (pNET) patients using only imaging data. These methods are described in detail in the hand-selected radiomic features and principal component (PC) radiomic features sections below. Features were extracted using the public PyRadiomics package. ¹² In addition, 86 Local Binary Pattern (LBP) and Local Ternary Pattern (LTP) features were extracted based on work by Gore S. et al. ¹³ Additionally, one method was used to compare normal pancreatic tissue to tumor tissue with contours and radiomic features from Radiomics Platform B. Normal pancreatic tissue contours were not available for Radiomics Platform A. This method is described in more detail in the Normal Tissue vs Tumor Tissue comparison section below.

Hand-Selected radiomic features

A small subset of hand-selected features were individually examined to identify any correlations with patient cancer type. These were total tumor volume, Mean HU, the 10th HU Percentile, and the 90th HU Percentile. Tumor volume was selected because of its ease of interpretability, and its direct correlation with the tumor's physicality. ¹⁴ The three intensity features were selected due to the well-established differences in enhancement between pancreatic adenocarcinoma and pNET patients. ¹⁵

Principal component radiomic features

Principal component analysis (PCA) was applied to entire feature groups and the top two PCs from each feature group (Table 1) were selected as composite radiomics features. We limited ourselves to the top two features per group to reduce the number of PC features examined, while also retaining the majority of the variance observed within each radiomics platform's entire set of features. The resulting eight PC features from Radiomics Platform A and ten PC features from Radiomics Platform B were analyzed for correlations with pancreatic cancer type. Radiomics Platform A provided no local texture features. Applying PCA reduced input feature sets from both radiomics Platforms A and B from 455 and 2078 features, respectively, to eight and ten PC features, respectively. If PC radiomic features were significant, the top contributing features were further analyzed to gain insights into what the PC feature measured.

Table 1.

Results of each selected and principal component (PC) radiomics feature's ability to separate pancreatic ductal adenocarcinoma (PDAC) patients from pancreatic neuroendocrine tumor (pNET) patients, as well as some basic statistics of the feature's values. HU = Hounsfield unit; AU = Arbitrary unit; AUC = Area under the curve; PC = Principal component; LBP = Local Binary Pattern; LTP = Local Ternary Pattern.

Normal tissue vs tumor tissue comparison

In addition to the above outlined tests, normal pancreatic tissue radiomics were available from Radiomics Platform B. A search for radiomic biomarkers to distinguish normal pancreatic tissue from cancerous pancreatic tissue was performed by examining a second set of hand-selected features. The Mean HU, 10th HU Percentile and 90th HU Percentile were selected to search for differences in tissue enhancement. Additionally, GLCM sum entropy was examined based on prior findings which demonstrated significant separation of tissue types in pancreatic adenocarcinoma patients.^16,17

Statistics

All statistical analyses were performed using SPSS, version 27 (IBM, Armonk, NY, USA), and p-values less than 0.05 were considered significant. All tests used were 2-tailed. Categorical data are expressed as frequencies and percentages, whereas continuous data are expressed as medians with interquartile range (IQR) or means with standard deviation where appropriate. Comparison of categorical variables among groups was performed using Chi-square or Fisher exact tests. Comparison of ordinal variables between groups was performed using Mann-Whitney U tests, as appropriate. PCA was used for dimensionality reduction. All radiomics analysis results were evaluated using Mann-Whitney U Rank Sum tests and their p-values along with area under the Receiver Operating Characteristic curve (AUC) value. To address the risk of inflated type I error from multiple hypothesis testing, we applied appropriate corrections for multiple comparisons. When multiple related outcomes were analyzed simultaneously, p-values were adjusted using a multiplicity control method (e.g., Bonferroni or false discovery rate), and adjusted p-values are reported where applicable. For exploratory analyses, results are interpreted cautiously with emphasis on effect sizes and confidence intervals rather than nominal statistical significance.

Results

A total of 25 patients were included from the study period. The mean age was 67.7 years (SD), and 68.0% were male. Of the included patient imaging, 20 patients had a histologic diagnosis of PDAC and five had pathology consistent with pNET. Neoadjuvant chemotherapy was administered in 80% of PDAC patients. A majority (60%) of patients underwent pancreaticoduodenectomy, and the remaining 40% had distal pancreatic resections. A high delta was present in 30% of PDAC patients.

Radiomics Platform A utilized the full 25-patient cohort, all with viable tumor contours. After Radiomics Platform B's annotation and feature generation, a total 23 patients had viable tumor contours while 24 had viable pancreatic contours. One PDAC and one pNET patient had boundaries of the pancreatic tumor which could not be confidently identified by either radiologist. In total, 19 PDAC and 4 pNET patients were available for analysis by Radiomics Platform B. One PDAC patient had a tumor region so large there was no viable normal pancreatic tissue remaining after subtracting the tumor region, leaving 24 patients with viable normal pancreas ROI.

No selected individual radiomics feature from either package showed discrimination between PDAC and pNET. When the top 2 PCA features from each of the four feature groups from Radiomics Platform A were examined, one of eight offered significant separation between PDAC and pNET (AUC=0.81, p = 0.038; Figure 2(a)). Interestingly, the five largest contributing factors to this component were all different measures of the tumor ROI's sphericity: sphericity, IBSI compactness1, asphericity, spherical disproportion, and compactness.

Figure 2.

Violin and receiver operating characteristic (ROC) plots for radiomics platform A and radiomics platform B’s most significant features are shown in subfigures 2a and 2b, respectively. Radiomics Platform A's shape principal component (PC) 2 feature separates the cohort's pancreatic neuroendocrine tumor (pNET) and pancreatic ductal adenocarcinoma (PDAC) patients with a p-value of 0.038 and an area under the curve (AUC) of 0.81, as shown in subfigure 2a. Radiomics Platform B's shape PC 2 feature separated the cohort's pNET and PDAC patients with a p-value of 0.017 and an AUC of 0.89, as shown in subfigure 2b. The physical interpretation of both Platforms’ shape PC 2 features is that lower values correspond to smoother (lower surface area to volume ratio), rounder (higher sphericity and lower flatness) tumors. AU = Arbitrary unit.

When the top 2 PCA features from each of the five feature groups from Radiomics Platform B were examined, a principal component of the second shape PC feature was identified as a significant discriminator of PDAC from pNET(Table 2). This feature had good predictive value overall (AUC=0.89, p = 0.019; Figure 2(b)).

Table 2.

Results of each selected radiomics feature's ability to separate normal pancreatic tissue from cancerous tumor tissue, as well as some basic statistics of the feature's values. HU = Hounsfield unit; AU = Arbitrary unit; AUC = Area under the curve; GLCM = Gray-level co-occurrence matrix.

Lower values of both features corresponded to smoother (lower surface area to volume ratio) and rounder (higher sphericity, lower flatness) tumors. This Implies PDAC tumors are more likely to be irregular in shape.

When tumor tissue was compared to normal pancreatic parenchyma, the GLCM sum entropy radiomics feature was observed to differentiate PDAC from normal gland (AUC=0.82, p < 0.0005; Figure 3(d)).

Figure 3.

Violin and receiver operating characteristic (ROC) plots for all selected radiomics features used to compare tumor and normal tissue. Note, this analysis was only possible with Radiomics Platform B's data and annotations. Subfigure 4a shows the plots for the region of interest's (ROI's) mean density, with nonsignificant results. Subfigure 4b shows the plots for the ROI's 10th density percentile, with a p-value of 0.003 and an area under the curve (AUC) of 0.76. Subfigure 4c shows plots for the ROI's 90th density percentile, with nonsignificant results. Subfigure 4d shows the ROI's gray-level co-occurrence matrix (GLCM) sum entropy, with a p-value of 0.00015 and an AUC of 0.82. HU = Hounsfield unit; AU = Arbitrary unit.

A subsequent analysis of the Platform B radiomics features revealed no evidence of confounding bias based on the contouring study reviewer or the original volume's contrast phase on downstream radiomic signals. This was tested by splitting the PDAC/pNET cohort features and Tumor/Normal tissue cohort features based on contouring study reviewer and contrast phase. The resulting feature distributions were then tested to see if they significantly differed. When comparing contouring study reviewer only the subset of patients contoured by both reviewers were included. All distributions for relevant radiomics features were found to be statistically consistent across both contouring study reviewer and volume contrast phase (Figure 4). For pNET patients, though distributions were found to be statistically consistent, the number of patients is likely too few to draw meaningful conclusions.

Figure 4.

Violin plots for possible sources of bias in the analyzed cohort. Subfigures 4a. and 4c. show comparisons of different annotating image reviewer for significant features. Subfigures 4b. and 4d. show comparisons of features extracted from Arterial and Portal Venous contrast phase computed tomography (CT) images. None of these splits show statistically significant separation between sub cohorts. AU = Arbitrary unit.

Of the 20 pancreatic adenocarcinomas studied, 6 (30%) were determined to be low-delta tumors. The average age of the low-delta tumor patients was 72.8 versus 67.1 years in high-delta (p = 0.035). The low-delta group was 50% male and 17% African American versus 64% and 0% in the high-delta group, respectively. A pancreaticoduodenectomy was performed in 67% of low-delta tumors and 57.1% of those with high-delta. The median BMI for the entire cohort was 26 (range 19.5 to 40.8). The mean BMI was 28.3 kg/m² for the low-delta group vs. 28.5 kg/m² for the high-delta group. The average tumor size was similar among the low-delta patients and the high-delta patients, at 3.1 cm vs. 2.7 cm, respectively (p = 0.486). All patients received neoadjuvant chemotherapy with either modified FOLFIRINOX or gemcitabine with nab-paclitaxel. By the end of the follow-up period, none of the low-delta patients experienced a recurrence event compared with 23.1% in the high-delta cohort.

Discussion

The results of this pilot study suggest that radiomics and radiologic features are a useful adjunct in the diagnosis and prognosis of pancreatic malignancy. Radiomics components were developed with highly specific discrimination between adenocarcinoma and other pancreatic tumors or normal pancreatic tissue. Our findings serve as a proof of concept that these image-generated data points can offer diagnostic value, even with a relatively small sample size. With further refinement of these techniques in larger datasets, we believe that radiomic-derived factors could become a reliable and noninvasive diagnostic tool for PDAC.

Pancreatic cancer is inherently more difficult to analyze via radiomics analysis than other previously studied tumors for several reasons. First, the ill-defined growth pattern of PDAC often makes annotation of tumors challenging and potentially subjective. Additionally, extrapancreatic extension of PDAC often is ill-defined and ambiguous for selection. The numerous vessels that border the pancreas also add complexity to the contouring process used to select the tumor for radiomics analysis. Another technical challenge is the frequent use of biliary stenting for pancreatic head cancers prior to high-quality imaging, which adds scatter artifacts to the images that may influence results. With these many obstacles, the automatic, algorithm-driven approaches to tumor segmentation utilized in lung and liver tumors are impracticable. Instead, the vastly more labor-intensive process of manual annotation of tumors is necessary for now in pancreatic tumors. With continual improvements in computer vision techniques, automated approaches to pancreatic segmentation that are robust and accessible are likely forthcoming.

As described in prior studies, delta was prognostic for recurrence in this group of cancer patients.^9,18 Recurrence events occurred more frequently with high-delta tumors than their low-delta counterparts. While clearly prognostic, delta is a radiologic feature that is limited by its dichotomous nature and subjectivity. Determining the set of radiomic data points that underlie a high- and low-delta is an ongoing effort but may help to demonstrate that radiomics features are also prognostic in PDAC. Radiomics has the advantage of objective, numeric data points that are continuous variables. Thus, the prognostic forecasting of radiomics ultimately may be of a more precise and reproducible nature.

This study has several limitations. Like all pilot studies, the sample size is limited to a small preliminary group of patients. The single-institution design offers consistent imaging protocols, but may restrict the overall statistical power and generalizability to all practice situations. The study is a retrospective by nature and has a nonrandomized design. Although a sequential inclusion methodology can limit selection bias to some extent, it cannot be completely eliminated. In addition, these software packages are novel and have mainly been geared toward alternative disease sites with homogenous background parenchyma. Extrapolating them for other organs carries some risk of bias from overrepresentation of certain factors. Furthermore, certain packages are proprietary, and it becomes impossible to determine how features were generated.

Quality statistical methods are necessary to interpret such large quantities of data. PCA and clustering can help with reduction analysis, but it is imperative to have appropriate guidance in bioinformatics. Frequent checks for sources of unseen bias are another important consideration. In the current study, interrater variability and contrast phase variability were minimal, but this is not always the case. Previous authors have published on the contouring variability seen with the irregular nature of pancreatic tumors. ¹⁹ Currently radiomics is not approved or accepted in diagnosing pancreatic cancer. Multiple studies have demonstrated is utilization for evaluation of pancreatic cysts, but it has not been adopted in the radiologic or surgical community at this time. Optimal processes still need to be refined as radiomics in pancreatic cancer becomes the subject of more investigation. We have demonstrated that radiomics could help bridge the diagnostic gap left by inconclusive biopsy by non-invasively characterizing tumor heterogeneity, differentiating malignant from benign mimics, inferring biological aggressiveness, and supporting multidisciplinary treatment decisions—particularly when tissue sampling is limited or unreliable.

Conclusion

In conclusion, we found that radiologic and radiomics features show promise as a reliable means to diagnose PDAC and are associated with clinical outcomes. Not all modalities used to extract these features are equivalent, though, and sound statistical methodology with organized bioinformatics are essential for clinical utility. Larger prospective evaluations including chronic pancreatitis patients are needed to underpin the accuracy and reliability of such novel techniques, but these early results are promising. An improved understanding of the radiomics features in pancreatic cancer may have implications for the assessment of prognosis, margin status, and treatment response.

Supplemental Material

sj-docx-1-cbm-10.1177_18758592261455545 - Supplemental material for Evaluation of radiomics and radiologic signatures for diagnosis and clinical outcomes in pancreatic adenocarcinoma and pancreatic neuroendocrine: A pilot study

Supplemental material, sj-docx-1-cbm-10.1177_18758592261455545 for Evaluation of radiomics and radiologic signatures for diagnosis and clinical outcomes in pancreatic adenocarcinoma and pancreatic neuroendocrine: A pilot study by Robert CG Martin II, Marc W Fromer, Jacob WH Gordon, Charles R Scoggins, Michael E Egger, Kelly M McMasters and Prejesh Philips in Cancer Biomarkers