Radiomics to Predict Microvascular Invasion in Hepatocellular Carcinoma: A Promising Biomarker for Tumor Recurrence

See also the article by Xia et al in this issue.

Dr Summers is a tenured senior investigator and staff radiologist at the NIH Clinical Center. His research interests include large radiology image databases, machine learning, and abdominal imaging. He is a fellow of the Society of Abdominal Radiologists and of the American Institute for Medical and Biological Engineering, a member of several journal editorial boards, and a recipient of the Presidential Early Career Award for Scientists and Engineers.

Primary liver cancer, of which hepatocellular carcinoma (HCC) comprises 75%–85% of cases, is the sixth most common malignancy and the third leading cause of cancer death worldwide (1). HCC has a poor prognosis and a high rate of recurrence after therapy. The main risk factors for HCC include chronic infection with hepatitis B or C virus, aflatoxin contamination in food, heavy alcohol consumption, obesity, type 2 diabetes, and smoking (1). Treatments include surgical resection, chemotherapy (transarterial chemoembolization), molecular therapy, immunotherapy, and liver transplant.

Abdominal CT is a mainstay of HCC diagnosis. CT is typically performed using a multiphasic technique that includes noncontrast, arterial, portal venous, and delayed phase imaging. The Liver Imaging Reporting and Data System is widely used to predict the likelihood of HCC on a multiphasic CT scan and requires a multiphasic technique (2). However, CT currently has a limited role in predicting the likelihood of HCC recurrence.

The article by Xia et al in this issue of Radiology (3) focuses on predicting microvascular invasion (MVI) in HCC on multiphasic abdominal CT scans. Some studies link MVI, the presence of tumor cells within the vasculature, to poor prognosis. For example, MVI is associated with metastatic disease (since vascular invasion by tumor permits hematologic spread) (4) and early recurrence (5). Unfortunately, the value of MVI as a risk predictor is somewhat limited, as it is generally best determined on the hepatectomy specimen and cannot be reliably determined from biopsies. However, a previous study has suggested that predicting MVI from imaging results could improve treatment planning for HCC (6).

In this retrospective study, the authors performed a radiomics analysis of the CT scans of patients with pathologically proven HCC. The study sample consisted of 773 patients, of whom 586 (76%) had hepatitis B virus infection. The presence of MVI was determined with histopathologic analysis of hepatectomy specimens. A radiologist manually delineated the tumors on the portal venous phase images. Radiomics features were extracted from the delineated tumors. Additional radiomics features were computed from subtraction images created from pairwise combinations of the noncontrast, arterial, portal venous, and delayed phase images. Additionally, three radiologists independently evaluated the tumors for five radiologic features: pseudo-capsule, two-trait predictor of venous invasion (TTPVI), peritumoral enhancement, tumor margin, and maximum tumor diameter. The feature TTPVI was defined as the presence of internal arteries within a tumor but without a hypodense halo around the tumor (7).

The patients underwent scanning at four medical centers in East Asia. The first of the four data sets became the training set used to train five radiomics models with logistic regression after feature reduction. The models were then tested internally on scans from the first medical center and then externally on scans from the second and third medical centers. The authors used scans from the fourth medical center as an outcome cohort to predict recurrence-free survival and overall survival. A fifth data set from the National Institutes of Health Cancer Imaging Archive was used for a small gene expression analysis.

The authors evaluated two models, one using radiomics alone and the other using a hybrid of radiomics features and radiologist-extracted features. Of the five radiomics models, the one that performed best on the internal set (the total model) incorporated 17 radiomics features. Of note, the total model did not perform significantly better than the other four models on the external test set. The hybrid model included features that performed well in a multivariable analysis, with odds ratios ranging from 2.23 to 6.78. In order from highest to lowest odds of predicting MVI, the features were TTPVI, pseudocapsule, peritumoral enhancement, and radiomics score. The authors found areas under the receiver operating characteristic curve ranging from 0.76 to 0.86 for the internal test set and from 0.72 to 0.84 for the external test set, respectively, depending on the model.

In a Kaplan-Meier analysis, the authors found that recurrence-free survival and overall survival could be categorized using both the total radiomics and hybrid models. With a median follow-up interval of 20.6 months, recurrence-free survival at 1 and 2 years was 74.4% and 71.1%, respectively; overall survival at 1, 3, and 5 years was 93.4%, 81.0%, and 73.6%, respectively.

In the fifth data set from The Cancer Imaging Archive, the authors found that in patients with MVI-positive tumors, differentially expressed genes showed involvement in glucose metabolism; those tumors also had low immune cell infiltration. It is unsurprising that tumors would be highly metabolically active with increased energy requirements and glucose metabolism, components of the Warburg effect. What is of greater interest is the reduced infiltration of immune cells in MVI-positive tumors; this finding suggests an approach to improving therapy based on greater immune cell activation or tumor penetration.

This is a comprehensive study of the radiologic and radiomics features of a reasonably large number of tumors from patients seen at four different medical centers. The study was carefully done, including standardization of radiomics features and external validations. At the same time, the study is quite complex, involving numerous experimental and statistical methods detailed in seven appendices.

The subtraction images and between-phase feature differences are a type of delta radiomics (8). Delta radiomics is a loosely defined term that refers to the calculation of differences between radiomics features as a function of time or contrast phase. Previously, delta radiomics have been successfully applied to CT scans of patients with HCC (9). In the article by Xia et al, the delta radiomics features were computed from different contrast phases of a single CT examination (3). Other delta radiomics approaches include feature computation from scans acquired at different stages of treatment. The use of delta radiomics is still in its infancy and requires further research.

Despite the attention to detail, this study had some limitations. Radiomics studies in general suffer from a dependency of the features on such factors as image noise and different acquisition parameters. The authors performed a small feature reproducibility study and selected only features that were reproducible. Nevertheless, the results need confirmation using the same method in a different patient sample. The hybrid method with the incorporation of radiologic features is subject to inter- and intrareader variability. A majority vote settled differences among the three readers. However, the authors did not report reader variability, which could be a serious issue given the potential subtlety of some of the radiologic features. A model that incorporated just the five radiologic features was not presented for comparison. Therefore, the added benefit of the radiomics features could be small. Most of the patients (76%) had hepatitis B virus–related HCC. Further validation is necessary before the method can be generalized to patients with other risk factors for HCC. The association of the imaging features with gene expression may not be generalizable because of the small sample size and limited information about the patients in the public data set used in the study.

If confirmed by others, the ability of radiomics to predict MVI in HCC could improve treatment planning (6). Potential applications impacting clinical management include guiding the size of the surgical margins and deciding whether to perform anatomic resection (hepatic segmentectomy) or transarterial chemoembolization. It would also be possible to identify which patients should undergo liver transplant versus surgical resection of the tumor. The results of this study suggest a potentially bright future for CT in the preoperative risk assessment in patients with HCC.