Overall survival in patients with HCC has remarkably improved owing to the recent advances in imaging diagnosis, surgical techniques, targeted therapies, and immunotherapy[20–24]. Despite these advancements, patients diagnosed with HCC and MVI for the first time present with low 5-year survival rates. MVI is found only in malignant tumors, as confirmed via postoperative pathology, with incidence ranging from 15.0–57.1%[3]. In this study, patient characteristics indicated that individuals with a history of cirrhosis and higher levels of AFP had a greater risk of MVI, which is consistent with previous studies[25–27].
Although conventional CT signs such as nonsmooth boundaries, hypohalos, and internal arteries may suggest the presence of MVI, the diagnostic accuracy relies on image quality and expertise of the radiologist. MVI prediction based on deep learning methods is considered a reliable method[10, 28–30], and such methods are being increasingly well-established; however, clear interpretable quantitative metrics are still required. Recent studies have shown that quantitative IC plays an important role in tumor diagnosis. Fan et al. closely reviewed relevant literature published in the past decade to explore the clinical value of DECT in the context of HCCs and concluded that DECT along with standardized quantitative parameters can be an extremely useful tool for HCC surveillance[31]. Marcon et al. reported that IC analysis using DECT can reliably distinguish papillary from clear cell subtypes of renal cell cancer, with a cutoff value of 3.1 mg/mL and accuracy of 86.8%[32]. DECT IC analysis revealed that the MVI-positive group had lower NICa and NICp values than the MVI-negative group (P < 0.05). These findings provide important insights into preoperative MVI risk assessment. However, the present study results contradict those reported by Kim et al.[33], who revealed that HCCs with MVI had significantly higher NICs than those without MVI (P < 0.05). Moreover, Yang et al. suggested that the MVI-positive group had a higher incidence of necrosis[34], potentially resulting in lower NICa and NICp values. These contradictory findings may be attributed to the differences in the sampling design, which require further investigation.
Most previous studies have chosen to use supervised learning for modeling, which may lead to risks of data dimensionality disaster, low model calculation efficiency, and overfitting. In order to solve this problem and help discover hidden structures and patterns in the data and help understand the data, we try to use PCA in unsupervised learning to solve these problems. PCA can retain the original data information to the greatest extent to achieve data dimensionality reduction, while outputting a linear combination of the original features. This combination ensures that the selected principal components can effectively represent the original data, which is helpful for improving the generalization ability of the model. and robustness are beneficial.
Utilizing PCA, we effectively reduced data dimensionality, discerning four principal components comprising differently weighted feature combinations. Collectively, these components explained 67.9% of the variance in the original dataset. Various features in the original data undergo reweighting based on the distinct weights assigned by these four principal components, resulting in the derivation of the corresponding principal component scores. The results revealed significant differences (P < 0.05) in PC3 and PC4 across Microvascular Invasion (MVI) groups in both the primary and validation datasets. Furthermore, PC3 and PC4 exhibited superior performance in MVI classification, with AUC values of 0.8410 and 0.8373, respectively. Moreover, bilirubin (TBIL, DBIL, and IBIL) was significantly correlated with PC3 (with factor loading weights of 0.9479, 0.8919, and 0.8849, respectively). Elevated bilirubin levels are usually indicative of abnormal liver function, and as liver cancer progresses, bilirubin dysregulation occurs. Also, ICa, Ica/ICaliver, and ΔICa were significantly correlated with PC4 (with factor loading weights of 0.8853, 0.8696, and 0.5607, respectively). The major characteristic of typical HCC, as defined by the Liver Imaging Reporting and Data System criteria, is the nonrim hyperenhancement in the arterial phase compared with normal liver tissue. Thus, ICa, Ica/ICaliver, and ΔICa were correlated with MVI in patients with HCC. This underscores the significance of these two distinct feature combinations in predicting MVI, aligning with findings reported by other scholars. Carr et al. analyzed a cohort of 2,416 patients with HCC and found that patients with higher bilirubin levels had poorer prognosis and indices of aggressiveness, confirming the significance of bilirubin in HCC[35]. Chan et al. identified high albumin bilirubin grading as a key parameter associated with early recurrence and developed a relevant preoperative model to predict the risk of early recurrence after HCC surgical resection[36].
This study has several limitations. First, this retrospective study involved two centers and lacked random data splitting for training and validation owing to the small sample size. Moreover, the two centers used different DECT imaging systems, which may have affected the results (center 1: dual tubes with beam filtration, center 2: rapid voltage switching with a single tube). However, regardless of the technology used for data acquisition, all commercial multienergy CT systems provide DECT data[37, 38]. Therefore prospective studies with a larger sample size should be conducted in the future, possibly including more types of DECT devices. Second, instead of volumetric iodine quantification, two-dimensional ROIs were used in this study; therefore, it is necessary to confirm the difference between the two measurements. Finally, morphological parameters were not included in our study. Further investigation is warranted to determine the potential benefits of combining quantitative data with morphological parameters for the preoperative prediction of MVI.