Quantitative imaging predicts microvascular invasion in hepatocellular carcinoma

Last Updated: 2017-09-28

By Will Boggs MD

NEW YORK (Reuters Health) - Quantitative image analysis of preoperative CT scans can be used to predict microvascular invasion in patients with hepatocellular carcinoma (HCC), researchers report.

"We used machine-learning techniques to extract information from routine CT scans for the detection of microvascular invasion (MVI)," Dr. Amber L. Simpson from Memorial Sloan Kettering Cancer Center, New York, told Reuters Health by email. "Knowing whether MVI is present in hepatocellular carcinoma before liver transplantation can potentially alter the way we allocate a limited resource - donated livers - to those patients who need it the most."

MVI is a strong predictor of recurrent disease after liver transplant, but so far it can only be diagnosed reliably in postoperative specimens.

Dr. Simpson and colleagues used preoperative CT quantitative image analysis to identify predictors of MVI in resected HCC tumors in a retrospective study of 120 patients who had undergone resection of HCC with curative intent. Their study was published online September 21 in the Journal of the American College of Surgeons.

They focused on two types of texture features: angle co-occurrence matrices (ACM), which quantify the orientation patterns of neighboring pixels over a specified distance and direction; and local binary patterns (LBP), which quantify the intensity patterns of the neighboring pixels.

Overall, 53 patients (44%) had tumors with pathologically confirmed MVI. Of patients whose tumors were 5 cm or less, 37% had MVI; of patients with tumors >5 cm, 49% had MVI. Recurrence-free survival was significantly worse when MVI was present, regardless of tumor size. Qualitative radiographic descriptors did not reliably predict MVI.

In contrast, 16 of 38 ACM features evaluated were significantly associated with MVI in tumors 5 cm or smaller, and 21 of 128 LBP features evaluated were significantly associated with MVI in tumors larger than 5 cm.

Using only the most significant features in multivariate analysis, a single ACM feature predicted MVI with 80% accuracy in patients with smaller tumors, and a single LBP feature predicted MVI with 75% accuracy in patients with larger tumors.

Adding in preoperative clinical data - alpha-fetoprotein level, largest tumor size, and history of viral hepatitis - further enhanced predictive value among larger tumors. For both categories of tumor size, negative predictive values could be optimized to yield 100% negative predictive value for tumors of all sizes.

"Our results underscore the potential for quantitative imaging to guide treatment of hepatocellular carcinoma, but we need to further test our models within a clinical trial setting," Dr. Simpson said.

"It is critically important for the academic community to find a way to share medical imaging data," she said. "With larger data sets, we can develop more-robust imaging biomarkers to alter clinical care in a meaningful way."

SOURCE: http://bit.ly/2wWJftZ

J Am Coll Surg 2017.