MRI imaging and machine learning based radiomics for detection of mixed HCC and CCA tumors

Abstract

Introduction and objectives: Primary liver cancer (PLC), comprising hepatocellular carcinoma (HCC) and cholangiocarcinoma (CCA), is a leading cause of cancer mortality globally. The combined hepatocellular-cholangiocarcinoma (cHCCCC) subtype may be less common but is relevant to treatment efficacy. We therefore evaluated the diagnostic accuracy of various approaches in distinguishing these liver cancers.

Materials and methods: Patients diagnosed with HCC, CCA, and cHCCCC at Beijing University Cancer Hospital and Institute, China were included. Radiologists of varying expertise independently assessed MRI scans, and we measured their diagnostic consistency. Radiomic features were extracted from MRI scans, and machine learning was applied to differentiate the cancer types.

Results: Standard imaging was insufficient to reliably characterize cHCCCC. Abdominal imaging experts (AIEs) had a higher mean sensitivity for HCC and CCA, 88% and 84% respectively, while non-experts (NIEs) had a lower sensitivity of 50% for HCC and 38% for CCA (HCC: p = 0.03, CCA: p = 0.008). Radiomic analysis found 'Sphericity' and 'ClusterShade' as the most relevant features. However, radiomics algorithms were also not sufficient to distinguish cHCCCC from either HCC or CCA. Regarding sensitivity, the radiomic-based model was not better than radiologists for any of the three classes (p = 0.065 for HCC, p = 0.426 for CCA, and p = 1.0 for cHCCCC). The random forest algorithm yielded an accuracy of 76% in the test set, since it correctly classified most HCC and CCA, while only one quarter of cHCCCC tumors.

Conclusions: Histopathological analysis, complemented by imaging as indicated, remains essential for accurate detection, diagnosis, and treatment of liver cancers.

Keywords: Diagnostic accuracy; Liver cancer; Machine learning models; Radiomics;MRI scans.