The differences in CT imaging seem to help us to accurately differentiate between cHCC-CCA and HCC. Previous studies have found that pseudocapsule, extensive necrosis (> 50%), the change of CT value in the arterial and venous phase, the change of CT value in arterial and delayed phase, targetoid appearances and non-smooth margin were helpful to distinguish cHCC-CCA and HCC [12, 13]. In our study, we found that density, morphology, hepatic lobe atrophy, capsule depression, bile duct dilatation, TIC, arterial phase rim enhancement, and peritumoral enhancement helped to identify cHCC-CCA from HCC. This is partially similar to the reports, but there are also inconsistent aspects, which are considered to be related to regional differences and tumor biological behavior. At the same time, we found that there were many imaging features overlapped in this study, which is similar to previous literatures [5, 14, 15]. Moreover, Huang et al. [12] found that some CT findings with statistical significance did not show their expected role in differential diagnosis after multiple regression analysis. Li et al. [16] demonstrated that the CT enhancement pattern of cHCC-CCA was similar to HCC or ICC. Therefore, it is difficult to accurately distinguish between cHCC-CCA and HCC only based on CT findings which is inefficient, and influenced easily by subjective factors, especially inexperienced radiologists [5].
In our study, there were some significant differences in serum biomarkers between cHCC-CCA and HCC, such as: elevated AFP and MVI were more common in HCC patients, while elevated CEA and CA-199, and cirrhosis were more common in cHCC-CCA patients, which are similar to the literature [2, 17, 18]. Recently, some researchers have combined serum biomarkers with CT contrast-enhanced patterns to improve the diagnostic accuracy of cHCC-CCA due to the lack of specificity in CT findings [5, 16]. Li et al [16] found that when elevated AFP and/or CA-199 inconsistent with the enhancement pattern on contrast-enhanced ultrasound (CEUS) or CT, it suggested a high possibility of cHCC-CCA, but the accuracy was only half. Zhou et al [14] proposed "CEUS LR-M and CT/MRI LR-M with AFP elevation" as the new diagnostic criteria for cHCC-CCA by combining the different diagnostic results of CEUS and CT/MRI enhancement pattern, tumour markers, and LI-RADS. Although the accuracy was improved, there was still nearly 30% misdiagnosis, and the sensitivity was low. In our study, the fusion model has better performance, with accuracy of 0.896, sensitivity of 0.914, specificity of 0.846 and AUC of 0.969. However, there were some abnormalities in the fusion model. The SHAP feature contribution of the 17 features showed that the SHAP values of the CT radiomics model were widely small, and the accuracy and AUC of the clinical model were slightly higher than those of the CT radiomics model, which may be due to the fact that the CHCC-CCA and HCC were similar in histopathology, leading to the decline of the diagnostic efficiency of the radiomics features. This may also be a reason why cHCC-CCA and HCC are not easily differentiated on CT findings [10].
Although the combination of CT findings, tumour markers and LI-RADS in the diagnosis of cHCC-CCA has attracted the attention, the role of radiomics combined with CT findings, clinical and serum biomarkers in the differentiation of cHCC-CCA from HCC may be ignored. Liu et al. [19] used support vector machine to evaluate the potential of MRI and CT radiomics features for predicting cHCC-CCA and non-cHCC-CCA, and found that CT radiomics features have limited diagnostic value in distinguishing between cHCC-CCA and non-cHCC-CCA. Among the AUCs of CT plain, arterial, venous, and delayed phases, the delayed-phase AUC was the best, but only 0.64. But that study has not combine clinical and serum biomarkers, so its diagnostic efficacy is lower than that of our study. Zhang et al. [10] developed a radiomics model to differentiate cHCC-CCA from ICC preoperatively, which consisted of CT arterial phase radiomics scores and clinico-radiological features, with an AUC of 0.942. Our results are similar to theirs, which suggests that a multi-factorial combined radiomics model is more effective in diagnosing cHCC-CCA.
We have some limitations. First of all, the sample size of cHCC-CCA is small, the main reason is the low incidence of cHCC-CCA and the center of research. Secondly, the tumor segmentation method used in this study is to select the largest layer and its adjacent upper and lower layers for segmentation, rather than the whole tumor segmentation method, which may not fully reflect the heterogeneity within tumors. However, some studies have shown that the performance of CT radiomics features obtained by 2D or non-whole tumor segmentation is similar to or even better than that of 3D [20–22]. Third, the model established in this study has not evolved external validation. The problem of sample size and single center can be solved by multi-center research, which can also increase the morphological imaging features of tumors to supplement the model with more details and complete external validation.
In summary, our study suggested that the fusion model composed of clinical factors, serum biomarkers, CT findings and CT radiomics features has good efficacy in distinguishing cHCC-CCA from HCC. These findings may provide a novel adjuvant approach for personalized clinical treatment strategies for patients with CHCC CCA.