Synopsis

The MRI-based Radiomics of pancreas can identify several imaging-characteristics (e.g. texture, shape, signal intensity, etc.) that are distinct in healthy and cancerous pancreas. We performed MRI-based radiomics of pancreas to demonstrate that radiomics play an important rule to differentiate healthy and cancerous pancreas and can assist diagnosis and management of PC. Multiple statistical tests demonstrated that 18% of the total 250 radiomic features were significantly different between healthy and cancerous pancreas. These features have high diagnostic accuracy to detect PC. We conclude that MRI-based radiomics of pancreas can potentially have a future role in early detection, prognosis, and prediction of treatment outcome of PC.

Introduction

Pancreatic cancer (PC) is the third leading cause of cancer-related death. It has a 5-year relative survival rate of 8% and is responsible for 7% of all cancer deaths (around 44,330 deaths) in the United States [1]. More than 80% of the patients are diagnosed with locally advanced PC. Clinicians often perform MRI examination of pancreas for early diagnosis and better management of PC. However, existing imaging biomarkers (such as pancreas inflammation, pancreatic ductal dilation, etc.) are often unclear and insufficient for early diagnosis, predicting treatment outcome, and determining likelihood of recurrence of PC. It is therefore essential to gain further insight into MRI characteristics of pancreas to improve differentiation between healthy and cancerous pancreas in order to have improved diagnosis and management of PC through MRI.

Radiomics [2, 3] is the process of extracting and analyzing hundreds of quantitative features from radiological images to identify distinct tissue patterns and regularities that might be imperceptible to human observation. With the assumption that several imaging-characteristics (e.g. texture, shape, signal intensity, etc.) are distinct in pancreas, we performed MRI-based radiomics of pancreas to demonstrate that radiomics can perform an important rule to accurately differentiate healthy and cancerous pancreas, in order to assist diagnosis and management of PC.

Material and Methods

The study included a cohort of 15 (10 male, 5 females, mean age = 65 years) and 30 (17 male and 13 females, mean age = 50 years) subjects with pancreatic cancer and healthy pancreas respectively. The T1-weighted fat-suppressed VIBE (Volumetric Interpolated Breath-hold Examination) images were acquired on a 3T system (Siemens Healthineers, Germany). The pancreas was manually segmented in all DICOM images (resolution: 2.3 mm in-plane and 3 mm in slice thickness) by a radiologist with 20-year experience and a clinician with 9-year experience. Note that the segmentation of cancerous pancreas contained both tumor and healthy part of the pancreas. Sample healthy and cancerous pancreas is shown in Figure 1.

An application was developed for automated extraction of radiomic features using Matlab-R2018. Signal intensities in all non-contrast MRI scans were discretized to multiple bins and normalized to unity, followed by extracting 250 common radiomic features from MRI images of both healthy and cancerous pancreas. The radiomic feature set for each group consists of 28 First-order statistics, 121 Gray level co-occurrence matrix statistics (Second-order statistics), 27 Gray level run length matrix statistics (Higher-order statistics), and 74 Geometry-based statistics.

The permutation test was carried out to compare the parameters of feature distributions of both groups using conditional Monte Carlo simulation with 5, 000 replicas. In addition, the Bonferroni-corrected method was used for multiple statistical comparisons of parameters. Moreover, the Bhattacharya coefficient was measured to examine the amount of overlap of distributions from two group.

Results and Discussion

A total of 250 radiomic parameters were extracted; out of which 18% (45/250) were found significantly different in PC and healthy pancreas (all P<0.0012 in permutation test), i.e. 6 calculated first-order statistics, 19 calculated co-occurrence matrix statistics, 8 calculated gray-level run-length matrix (GLRLM) parameters, and 12 calculated geometry-based parameters were significant (see Manhattan plot in Figure 2). In addition, no significant difference was observed between pancreatic tissue regarding patient characteristics and image quality parameters.

The Bhattacharya coefficient was used to measure the amount of overlap of distributions from two groups. The corresponding distributions of all features with significant difference in two groups (P<0.0012) passed a predefined overlap criterion, i.e. less than 0.3 Bhattacharya coefficient is considered distributions of significantly different features; whereas Bhattacharya coefficient 1 is considered 100% overlap of distributions.

Furthermore, to validate the results of radiomics, a Naïve Bayes probability classifier was used that utilized 5 randomly selected features (out of 45 identified radiomics features) to classify MRI scans into PC and healthy class. A five-fold cross-validation is performed using the same data that was used for radiomics. The overall classification accuracy achieved is 0.87. The result validates the discriminatory value of identified radiomic features.

Conclusion

Acknowledgements

References

[1] Cancer Facts and Figures, American Cancer Society, 2018 https://www.cancer.org/research/cancer-facts-statistics/all-cancer-facts-figures/cancer-facts-figures- 2018.html

[2] Gillies, Robert J; Kinahan, Paul E; Hricak, Hedvig (2016). "Radiomics: Images Are More than Pictures, They Are Data". Radiology. 278 (2): 563-577.

[3] Ranjbar, Sara; Ross Mitchell, J (2017). "An Introduction to Radiomics: An Evolving Cornerstone of Precision Medicine". Biomedical Texture Analysis. pp. 223-245.