Radiologists assess subtle changes in contrast enhancement patterns on imaging to determine diagnosis and treatment response. Radiomics is an emerging field that attempts to derive quantitative information from enhancement patterns in imaging (1,2). Similar to genomics and proteomics, radiomics has the potential to improve our understanding of cancer biology using quantitative methods. Figure 1 shows a standard diagnostic abdominal CT scan with a region of interest (ROI) magnified. Changes in pixel values across the magnified region can be appreciated. Texture analysis is a technique that describes the spatial variation of pixel intensity in an ROI where the ROI could be any biologically relevant structure such as parenchyma, tumor, or node. Extracting additional information from routine imaging is an attractive opportunity for clinicians looking for low cost methods to better inform patient management. Below we describe four projects to develop radiomic technologies with an eye toward the ultimate goal of precision therapy for hepatic and pancreatic tumors.
Prediction of Response to Chemotherapy
With support of the Society of Memorial Sloan Kettering, we are investigating whether quantitative imaging features extracted from routine computed tomography (CT) scans can be used to predict response to chemotherapy in patients with colorectal liver metastases (CRLM). There is no current method of predicting, prior to treatment, whether patients will respond to chemotherapy. Such a method would fundamentally alter our approach to cancer treatment. Based on the variable appearance of CRLM on contrast enhanced CT, our hypothesis is that quantitative imaging features can detect differences in tumor perfusion affecting chemotherapy delivery. Figure 2 illustrates two patients with CRLM: the patient with a marked volumetric response (A, B) had colorectal liver metastases with more heterogeneous enhancement before treatment (A), compared to the patient with a poor response (C).
In 2016, we reported preliminary findings on this study at the annual meeting of the Radiological Society of North America. We demonstrated that texture analysis of pre-treatment scans predicts volumetric response. 103 patients with unresectable CRLM from two prospective clinical trials studying combined systemic and hepatic arterial infusion (HAI) chemotherapy were included in the retrospective study. Index tumors were segmented from CT at baseline and after 8 weeks of HAI treatment. Volumetric response was assessed as the percentage change in tumor volume from baseline to follow-up. Imaging features (summary statistics, texture, and shape properties) were extracted from the tumor region in the baseline scan. Imaging features statistically significant on univariate analysis were included in a regression model. Data were randomly split into training (n=93) and test (n=10) sets. Random forest regression models were employed with cross-validation on the training set. Test data were entered into the trained regression models. Prediction error for the test data was 13.32% (CI: 12 – 14.65%). Figure 3 illustrates exemplar responder and non-responder tumors with the histogram of two features: Hounsfield units and short run emphasis (measures consecutive pixels with the same intensity values). These results support that increased heterogeneity may be related to better response and suggest that greater intravenous contrast uptake may translate to greater intake of chemotherapy into CRLM and subsequent volumetric response.
Pancreas Cancer Survival Prediction
With support from Cycle for Survival, we are evaluating tumor texture as a preoperative prognostic marker in pancreatic ductal adenocarcinoma (PDAC). PDAC is one of the most lethal cancers worldwide with a 5-year overall survival rate of 6 percent (3). Complete surgical resection, achievable in 10 percent to 15 percent of patients, is the only curative treatment. Therefore, determining preoperative prognostic factors is crucial for these patients.
We have demonstrated that extracting quantitative imaging features from pre-treatment scans can provide patient prognosis stratification. The Kaplan-Meier curve in Figure 4 demonstrates the power of one imaging features to stratify PDAC patients into distinct groups. In particular, imaging features combined with differentiation enabled identification of a group of patients with a 60 percent chance of survival at 5 years — a group previously unknown in the study of PDAC. This marker was independent of all other clinical variables. Although this proof of principle needs further validation across a sufficiently powered study, it nonetheless shows promise for the evaluation of patients with pancreatic cancer.
Liver Failure Prediction
Recently, we demonstrated that underlying liver insufficiency is correlated with heterogeneous hepatic parenchymal enhancement, which can be quantified with radiomics (4). The appearance of the liver parenchyma from preoperative CT images of patients with postoperative liver insufficiency was significantly different than that of patients with no postoperative liver insufficiency (Figure 5). Hepatic insufficiency has the potential to delay chemotherapy treatment, prolong hospital stay, and increase the overall risk of cancer recurrence, so the importance of identifying patients at risk is clear.
Radiogenomics of Cholangiocarcinoma
Radiogenomics is an emerging field focusing on establishing relationships between imaging features (radiomics) and molecular markers (genomics). Advances in radiogenomic imaging have the potential to contribute to clinical decision making through development of predictive and prognostic treatment algorithms and noninvasive disease surveillance. Standard biopsy requires invasive tissue procurement procedures that lack temporal and spatial dimensions, as they provide information in a single time point, typically from a single anatomical site. By contrast, the radiogenomic approach can be implemented in multiple time points and at multiple tumor sites. In a study of intrahepatic cholangiocarcinoma, an aggressive primary liver cancer, we demonstrated that imaging features correlated with EGFR and VEGF expression levels (Figure 6). Our results suggest that noninvasive radiogenomic methods may predict protein expression of intrahepatic cholangiocarcinoma (5).
- Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van Stiphout RG, Granton P, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer. 2012;48(4):441–6.
- Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology. 2015;278(2):151169.
- Siegel R, Naishadham D, Jemal A. Cancer statistics, 2012. CA Cancer J Clin. Jan;62(1):10–29.
- Simpson AL, Adams LB, Allen PJ, D’Angelica MI, DeMatteo RP, Fong Y, et al. Texture Analysis of Preoperative CT Images for Prediction of Postoperative Hepatic Insufficiency: A Preliminary Study. J Am Coll Surg. 2015;220(3):339–46.
- Sadot E, Simpson AL, Do RKG, Gonen M, Shia J, Allen PJ, et al. Cholangiocarcinoma: Correlation between Molecular Profiling and Imaging Phenotypes. PLoS One. 2015 Jan;10(7):e0132953.