Prediction of Histopathologic Growth Patterns of Colorectal Liver Metastases with a Noninvasive Imaging Method.

ABSTRACT

OBJECTIVES: To predict histopathologic growth patterns (HGPs) in colorectal liver metastases (CRLMs) with a noninvasive radiomics model. METHODS: Patients with chemotherapy-naive CRLMs who underwent abdominal contrast-enhanced multidetector CT (MDCT) followed by partial hepatectomy between January 2007 and January 2019 from two institutions were included in this retrospective study. Hematoxylin- and eosin-stained histopathologic sections of CRLMs were reviewed, with HGPs defined according to international consensus. Lesions were divided into training and validation datasets based on patients' sources. Radiomic features were extracted from pre- and post-contrast (arterial and portal venous) phase MDCT images, with review focusing on the segmented tumor-liver interface zones of CRLMs. Minimum redundancy maximum relevance and decision tree methods were used for radiomics modeling. Multivariable logistic regression analyses and ROC curves were used to assess the predictive performance of these models in predicting HGP types. RESULTS: A total of 126 CRLMs with histopathologic-demonstrated desmoplastic (n = 68) or replacement (n = 58) HGPs were assessed. The radiomics signature consisted of 20 features of each phase selected. The 3 phases fused radiomics signature demonstrated the best predictive performance in distinguishing between replacement and desmoplastic HGPs (AUCs of 0.926 and 0.939 in the training and external validation cohorts, respectively). The clinical-radiomics combined model showed good discrimination (C-indices of 0.941 and 0.833 in the training and external validation cohorts, respectively). CONCLUSIONS: A radiomics model derived from MDCT images may effectively predict the HGP of CRLMs, thus providing a basis for prognostic stratification and therapeutic decision-making.