Le Guo,1,2,* Chantao Huang,2,* Peng Hao,2,* Ningyang Jia,3 Ling Zhang2,4 

1Lab of Molecular Imaging and Medical Intelligence, Department of Radiology, Longgang Central Hospital of Shenzhen (Shenzhen Clinical Medical College, Guangzhou University of Chinese Medicine), Shenzhen, 518116, People’s Republic of China; 2Department of Imaging Diagnosis, Nanfang Hospital, Southern Medical University, Guangzhou, 510515, People’s Republic of China; 3Department of Radiology, Shanghai Eastern Hepatobiliary Surgery Hospital, Second Military Medical University, Shanghai, 200438, People’s Republic of China; 4Medical Imaging Department of Ganzhou People’s Hospital, Ganzhou, 341000, People’s Republic of China

*These authors contributed equally to this work

Correspondence: Ling Zhang, Department of Imaging Diagnosis, Nanfang Hospital, Southern Medical University, Guangzhou, 510515, People’s Republic of China, Tel +86-2062786561, Email zhangling_smu@163.com Ningyang Jia, Department of Radiology, Shanghai Eastern Hepatobiliary Surgery Hospital, Second Military Medical University, Shanghai, 200438, People’s Republic of China, Tel +86-02181875211, Email jianingyang6@163.com

Objective: To investigate the role of MRI peritumoral imaging in predicting microvascular invasion (MVI) status in patients with combined hepatocellular carcinoma and cholangiocarcinoma (cHCC-CCA).
Methods: Clinical and pathological data and MRI images of 118 patients with surgically resected and pathologically confirmed cHCC-CCA were retrospectively collected. The tumor in MRI images was segmented by ITK-SNAP software in three dimensions and extended 1 centimeter(cm) towards the tumor periphery. Then, the Python open-source platform was used for radiomics analysis. Mutual information and recursive elimination methods were used to select the optimal features. Clinical models and radiomics models were constructed based on six classifiers. The model’s effectiveness was comprehensively evaluated using receiver operating characteristic (ROC), area under curve (AUC), and decision curve analysis (DCA), and the model results were output using Shapley Additive exPlans (SHAP).
Results: The differences in HBeAg, capsule, target sign, and lymph node metastasis between MVI negative and positive groups were statistically significant (p < 0.05). Based on peritumoral, 1cm fusion model (in arterial phase) has an AUC of 0.940 (95% CI: 0.801– 0.947) and 0.825 (95% CI: 0.633– 0.917) in the training/testing set when identifying the MVI status of cHCC-CCA. The accuracy, sensitivity, and specificity in the testing set are 0.778, 0.800, and 0.726, respectively. The DCA shows that when the threshold is approximately 11.08%– 66.47%, the net return of the fusion model is higher than that of the clinical and radiomics models under the same conditions.
Conclusion: Radiomics with a 1cm extension around the tumor can improve the performance of machine-learning models in predicting MVI labels.

Keywords: peritumoral radiomics, microvascular invasion, combined hepatocellular carcinoma and cholangiocarcinoma, Shapley additive explans