Abstracts

RETURN TO ABSTRACT LISTING


2054. Accuracy in Radiomics Features vs. Radiologists in Differentiation of T1 Stage Pancreatic Neuroendocrine Tumor from Healthy Controls
Authors * Denotes Presenting Author
  1. Seyoun Park; Johns Hopkins University
  2. Satomi Kawamoto; Johns Hopkins University
  3. Sheila Sheth; Johns Hopkins University
  4. Javad Azadi; Johns Hopkins University
  5. Elliot Fishman; Johns Hopkins University
  6. Linda Chu *; Johns Hopkins University
Objective:
Pancreatic neuroendocrine tumors (PanNETs) have variable imaging appearances. Previous publications showed that radiomics features are helpful in differentiating tumor grades in PanNETs and in differentiating PanNETs from other pancreatic tumors. The purpose of this study is to determine whether radiomics features can be used to detect small T1 stage panNETs.

Materials and Methods:
As an IRB-approved matched case-control study, 138 cases (66 male + 72 female; 56 ± 14 years) of pathologically resected T1 stage PanNETs and 140 cases (65 male + 75 female; 55 ± 9 years) of healthy controls were retrospectively selected from the radiology databases from 2012-2017. All participants underwent IV contrast-enhanced CT with arterial and venous phase and the arterial phase were used for the classification. Whole 3D volume of the pancreas (including tumor and background pancreas) were manually segmented from arterial phase images (VelocityAI, Varian Medical Systems). 478 radiomics features were extracted, which included first-order statistics, shape, texture, and texture features from wavelet and Laplacian of Gaussian filtering. Cases were randomly split into training set (89 PanNETs + 93 controls) and testing set (49 PanNETs + 47 controls). A random forest was applied for PanNET classification. A radiologist reviewed all PanNET cases and measured attenuation of PanNETs relative to background pancreas and subjectively graded the visibility of PanNETs. Two other radiologists who were blinded to the diagnosis independently reviewed the test cases to evaluate for presence or absence of PanNETs.

Results:
Mean PanNET tumor size was 1.27 ± 0.35 cm. In the arterial phase, 70 PanNETs were hyperenhancing, 27 cases were isoenhancing, and 41 cases were hypoenhancing relative to background pancreas. On visual assessment of 138 cases of PanNETs, 71 PanNETs were visible, 44 cases were barely visible and 23 cases were not visible in the arterial phase. In test set of PanNET cases, 61.1% (30/49) were graded as visible, 20.4% (10/49) were barely visible, and 18.4% (9/49) were not visible on arterial phase. The overall radiomics-based classification accuracy of the test set was 87.5% (84/96), with PanNET sensitivity of 93.9% (46/49) and specificity of 80.1% (38/47). In comparison, the average accuracy of radiologists was 75.0%, with sensitivity of 51.02% and specificity of 100%.

Conclusion:
Radiomics features can accurately differentiate T1 stage pancreatic neuroendocrine tumor from normal pancreas and can achieve higher sensitivity and accuracy compared to radiologists. Radiomics has potential to be a helpful second-reader to improve sensitivity of pancreatic neuroendocrine tumor detection.