Unsupervised Machine Learning of MRI Radiomics Features Identifies Two Distinct Subgroups with Different Liver Function Reserve and Risks of Post-Hepatectomy Liver Failure in Patients with Hepatocellular Carcinoma

Abstract

Objective: To identify subgroups of patients with hepatocellular carcinoma (HCC) with different liver function reserves using an unsupervised machine-learning approach on the radiomics features from preoperative gadoxetic-acid-enhanced MRIs and to evaluate their association with the risk of post-hepatectomy liver failure (PHLF). Methods: Clinical data from 276 consecutive HCC patients who underwent liver resections between January 2017 and March 2019 were retrospectively collected. Radiomics features were extracted from the non-tumorous liver tissue at the gadoxetic-acid-enhanced hepatobiliary phase MRI. The reproducible and non-redundant features were selected for consensus clustering analysis to detect distinct subgroups. After that, clinical variables were compared between the identified subgroups to evaluate the clustering efficacy. The liver function reserve of the subgroups was compared and the correlations between the subgroups and PHLF, postoperative complications, and length of hospital stay were evaluated. Results: A total of 107 radiomics features were extracted and 37 were selected for unsupervised clustering analysis, which identified two distinct subgroups (138 patients in each subgroup). Compared with subgroup 1, subgroup 2 had significantly more patients with older age, albumin–bilirubin grades 2 and 3, a higher indocyanine green retention rate, and a lower indocyanine green plasma disappearance rate (all p < 0.05). Subgroup 2 was also associated with a higher risk of PHLF, postoperative complications, and longer hospital stays (>18 days) than that of subgroup 1, with an odds ratio of 2.83 (95% CI: 1.58–5.23), 2.41(95% CI: 1.15–5.35), and 2.14 (95% CI: 1.32–3.47), respectively. The odds ratio of our method was similar to the albumin–bilirubin grade for postoperative complications and length of hospital stay (2.41 vs. 2.29 and 2.14 vs. 2.16, respectively), but was inferior for PHLF (2.83 vs. 4.55). Conclusions: Based on the radiomics features of gadoxetic-acid-enhanced MRI, unsupervised clustering analysis identified two distinct subgroups with different liver function reserves and risks of PHLF in HCC patients. Future studies are required to validate our findings.

Machine Learning Combined with Radiomics Facilitating the Personal Treatment of Malignant Liver Tumors

“…In groups with indocyanine green retention rates at 15 min set at 10%, 20%, and 30%, the MRI radiomics model outperformed CT, with AUCs of 0.917 vs. 0.822, 0.979 vs. 0.860, and 0.961 vs. 0.938, respectively. Similarly, Wang and collaborators [51] employed preoperative gadoxetic acid-enhanced MRI radiomics features and an unsupervised machine learning approach to assess the risk of liver failure in HCC patients with varying functional liver reserves, revealing significant distinctions among functional liver reserve subgroups.…”