Page 157 - Read Online

Page 224 Grewal et al. Art Int Surg 2023;3:217-32 https://dx.doi.org/10.20517/ais.2023.28

comparison, a prognostic nomogram incorporating age, N stage, M stage, tumor size, and surgery yielded
[117]
AUCs of 0.909 and 0.890 for predicting 1-year and 5-year survival, respectively . Similar studies have
demonstrated certain radiomics features to accurately and independently predict post-surgical overall
survival [118,119] . Seeing as HCC recurrence rates range from 25% after liver transplantation to over 70% after
liver resection or RFA , numerous studies investigating the potential of radiomics in predicting
[120]
recurrence risk after various treatments are also underway. In a multi-institutional study, based on contrast-
enhanced computed tomography (CECT) analysis, preoperative and postoperative ML-based radiomics
models were developed to predict HCC recurrence after liver resection, and both exhibited better
prognostic performance compared to the current staging system (c-index = 0.733-0.801) . Another study
[121]
used CECT images from 184 patients to derive radiomics signatures of risk-straying patients based on their
[122]
recurrence-free survival after RFA (c-index = 0.755) . Other groups have also utilized DCE-MRI to
develop radiomics models; for instance, one study compared the prognostic performance of radiomic scores
for the tumor region, perilesional region, non-tumoral region, which independently had the highest
prognostic score (c-index = 0.72); the combined score of the three regions yielded a c-index of 0.83 .
[123]
Focusing on patients with unresectable disease, a separate study group used preoperative multiparameter
MRI (mp-MRI) to create a radiomics model to predict early progression after TACE (AUC: 0.800) .
[106]
Furthermore, studies suggest that models combining radiomic and clinical input may provide the best
performance. For instance, one study established a pretreatment CT-based radiomics model for overall
survival after TACE and reported that a composite model was more predictive than a radiomics or clinical
[106]
signature alone (HR: 19.88, 95%CI: 6.37-92.02, P < 0.001) .
Similar efforts to estimate treatment response and survival for CCA, which remains a highly lethal
hepatobiliary malignancy, have been conducted. A number of studies have radiomics-based models to
predict lymph node metastasis (LNM) in ICC patients with validation cohort AUCs ranging from 0.80 to
0.89 [124-126] . Two studies have developed radiomics-based models for preoperative prediction of
microvascular invasion (MVI) [127,128] . In another study focusing on LMN in ECC, an MRI-based radiomics
[129]
model achieved an AUC of 0.9 in predicting LNM . Other researchers have explored recurrence after
surgery with implications for informing surveillance schedules and any adjuvant therapy. For example, one
group employed random forest and logistic algorithms to develop clinical, radiomic, and combined models;
the radiomics model had higher AUCs (sensitivity 0.846, specificity 0.771 in the validation cohort) than the
clinical model and comparable AUCs to the combined model .
[100]
Altogether, current literature underscores that the incorporation of radiomics may enhance models to
predict response to various treatments, thereby guiding appropriate treatment selection and sequencing. In
facilitating more informed clinical decision-making, radiomics models will hopefully help improve patient
outcomes for hepatobiliary cancers.

CURRENT PITFALLS
Despite promising results, the field of radiomics is still in its relative infancy. As of now, there are still
certain critical limitations across the majority of literature on radiomics in hepatobiliary and pancreatic
diseases, which need to be addressed prior to its clinical integration [Table 1].

First, a vast majority of existing literature on radiomics comprises retrospective single-institution series with
relatively small sample sizes. Currently, a majority of these radiomics-based models have not been validated
in other datasets and are therefore not generalizable. Second, prior to the extraction of radiomic features
from images, the areas of interest within those images must be delineated, a process known as
segmentation . Current segmentation methods are either manual, semi-automated or fully automated.
[22]