CT texture features are associated with overall survival in pancreatic ductal adenocarcinoma – a quantitative analysis

Cancers are phenotypically heterogeneous and their pattern of spatial heterogeneity varies with time [1]. Tumor heterogeneity is thought to be a key factor in the development of therapeutic resistance [2]. Tumor genomics from needle biopsy may be reflective of only a portion of the cancer’s characteristics. However, imaging has the distinct advantage of being non-invasive and providing an overview of the entire tumor. Therefore, imaging has increasingly been used to capture spatial heterogeneity of tumors [3]. The in-depth feature analysis of tumor sites has been brought into the “omics” terminology, called radiomics, which is defined as the high-throughput extraction of image features from radiographic images [1, 4, 5]. Imaging features can be derived from standard of care modalities such as contrast-enhanced computed tomography (CT), magnetic resonance imaging (MRI) and positron emission tomography (PET) without modification of the acquisition protocols making them less cost prohibitive [6‐8]. For example, texture features from grey level co-occurrence matrices (GLCM) [9], which generate second-order statistical features have been used and improved [10] to quantify spatial texture of objects. There is an abundance of literature suggesting GLCM and other texture traits [11] are significantly associated with overall survival in lung [12], breast [13] and hepatic [14] carcinomas. However, to our best knowledge, there is paucity of studies to date about the potential prognostic value of CT texture features in pancreatic ductal adenocarcinoma (PDAC) [15].

PDAC has the lowest 5-year overall survival (OS) rate of any epithelial carcinoma at 7.7% [16] and surgical resection, applicable to < 30% cases [17], is the only potential cure [18] increasing OS to about 15–20% for resected cases [19]. More recently, neoadjuvant therapy has been introduced with the hope of extending survival for patients with resectable disease, allowing resection in patients with initially unresectable disease and selecting patients with different natural histories and chemosensitivities [20]. Since contrast-enhanced CT imaging is routinely used [21] for assessing resectability, staging and assessment of disease progression [22], a CT-derived quantitative imaging biomarker of OS could potentially provide a window into prognosis of PDAC.

The purpose of this study was to assess whether radiomic features from pre-operative contrast-enhanced CT in resectable PDAC patients were associated with overall survival. We hypothesized that pre-selected texture features are associated with the OS in resectable PDAC patients.

The demographic information of the cohort is shown in Table 2. Tumor region and the normal pancreas were used for analysis. The median (interquartile range) of HU was 71 (61–82) and 57 (41–63) in normal pancreas and tumor, respectively (p < 0.001). The HU was significantly higher in normal tissue than tumor regions in all patients (p < 0.001) for the portal venous phase. Tumor was significantly different than normal pancreas, as shown in Table 3; uniformity (p < 0.001), entropy (p = 0.009), and correlation (p < 0.001). However, the difference in dissimilarity (p = 0.530) and inverse difference normalized (p = 0.511) were not significant.

Wald-test in Cox regression analysis on tumor texture parameters showed dissimilarity (coefficient = −0.1292, p-value = 0.045) and inverse difference normalized (coefficient = 71.81, p-value = 0.046) were significantly associated with OS as shown in Table 4. Kaplan-Meier plots of cumulative survival for significant tumour features are provided in Fig. 2. Also, size of the tumor (coefficient = 0.000627, p-value = 0.611) and the average intensity of the tumor (coefficient = −0.011, p-value = 0.366) were not significantly associated with OS. Other tumor texture features were not significantly associated with survival with the coefficients and p-values as follows: uniformity (coefficient = −105.5, p-value = 0.334), entropy (coefficient = 0.324, p-value = 0.330), correlation (coefficient = 4.013, p-value = 0.068).

Among the studied features, dissimilarity (AUC = 0.716) and inverse difference normalized (AUC = 0.716) showed maximal predictive value for predicting OS. AUC for uniformity, entropy, and correlation were 0.560, 0.569, and 0.680, respectively. Table 5 represents details of the ROC analysis.

Figure 3 shows the histogram of two significant features namely dissimilarity and inverse difference normalized. This figure also illustrates the distribution of survival across the features values.

In our study, we used GLCM textural analysis from venous phase contrast-enhanced CT and found that dissimilarity and inverse difference normalized were associated with OS in a cohort of resectable PDAC patients. These features provided a stronger association with OS than tumor intensity and tumor size.

We found that less inverse difference normalized and greater dissimilarity are associated with longer OS. As texture analysis has not been used in the context of resectable PDAC, there were no studies to compare our findings. However, we can compare our results with other studies that followed a similar hypothesis in other types of cancer. Our findings are consistent with previous radiomic studies for lung, breast, and other cancer site. Different studies have also shown that dissimilarity feature extracted from ROIs in PET/CT has a positive correlation with survival time for non-small cell lung cancer [35, 36] and Multi–Cancer site patient cohorts [37]. It is important to note that the underlying meaning of imaging texture cannot be reduced to a mean regional intensity but is estimated from the probability of the occurrence of a specific pattern of intensities which can explain why mean intensity was not associated with survival outcome in our study and others. Such texture features may not be readily visible on standard grey scale images.

Given the challenging management of PDAC which presents late and is highly lethal, it is hoped that gathering clues of factors associated with overall survival from CT could augment the ability for treatment decision making [38] and aid in prognosticating treatment scenarios. Our findings show that with minimal cost and with no additional imaging burden, textural feature analysis of routine contrast-enhanced CT imaging before surgery may provide useful information for PDAC patients undergoing curative intent surgical resection. Our results show that textural analysis is more strongly associated with OS than tumor size. Such informed decision may help in identifying therapeutic plan for patients, for example, considering targeted adjuvant or neoadjuvant treatments in some patients with predicted poor prognosis. On the other hand, it may help in identifying patients with very poor prognosis who are unlikely to benefit from surgery; in these patients chemotherapy or radiation may be the optimal treatment modality [39]. Certainly for clinical application in personalized medicine, a wider repertoire of treatment options with better survival is a fundamental challenge in this nearly uniformly lethal disease. As our understanding of imaging biomarkers continues to unfold, it is hoped that this may provide more insight into the likely benefit of new therapeutic regimes in subpopulations of patients.

This study has limitations. The small sample size limits our ability to perform a multivariate analysis and thus additional stratification based on tumor extent such as resectable versus borderline resectable stratification could not be evaluated. Despite a relatively small sample size, we found strong association of the textural features with OS. The findings of this study encourage investigating the association of a wider range of radiomic features with survival and intermediate factors such as radiogenomics in a larger sample size. Also the reported findings should be verified independently in future studies. Further work is also needed to address the repeatability of these quantitative imaging biomarkers as part of a biomarker validation process [40]. There is always a risk of achieving statistically positive results by chance with small sample sizes. We have tried to restrict the number of features being evaluated by choosing features that have already been shown to be prognostic in a variety of other adenocarcinomas [12, 26]. The fact that some of the texture features associated with OS match those in other adenocarcinomas is encouraging. Finally, there is a lack of understanding on the underlying relationship of texture features and histology, genomics and proteomics of PDAC which requires further work.

CT-derived PDAC texture features of dissimilarity and inverse difference normalized are promising prognostic imaging biomarkers of OS for patients undergoing curative intent surgical resection.