Page 170 - Read Online

Endo et al. Art Int Surg 2024;4:59-67 https://dx.doi.org/10.20517/ais.2024.09 Page 63

Radiomics has emerged as a tool to diagnose, risk stratify, and predict prognosis among patients with
HCC [45,46] . For instance, several studies have focused on developing preoperative models to predict
microvascular invasion, an important prognostic factor that is traditionally identified only after surgery on
pathological examination [47,48] . Jiang et al. demonstrated the effectiveness of a ML-based model combining
radiomics with other clinicopathologic factors that resulted in high predictive accuracy . In the future,
[49]
radiomics will likely play a significant role in facilitating decision making regarding the treatment of HCC.

Biliary tract cancer
[50]
Bile duct cancer is a rare disease that often poses a diagnostic challenge . To assist in tumor classification
and postsurgical prognosis, several researchers have used ML-based approaches to the analysis of
international multi-institutional datasets to improve the performance of prognostic tools. For instance,
Alaimo et al. developed and validated three ML models aimed at predicting early recurrence (within
< 12 months after hepatectomy) . Notably, the RF model demonstrated the highest discrimination with an
[51]
AUC of 0.904 in the training cohort and 0.779 in the validation cohort. The top five influential variables
were the tumor burden score , perineural invasion, microvascular invasion, carbohydrate antigen
[52]
(CA)19-9, and nodal status. In a different study by Cotter et al., a classification and regression tree (CART)
approach was employed to stratify gallbladder carcinoma (GBC) patients relative to OS . Interestingly,
[53]
CART analysis identified tumor size, biliary drainage, CA19-9 levels, and the neutrophil-lymphocyte ratio
(NLR) as the factors most strongly associated with OS, effectively classifying patients into four prognostic
groups. Tsilimigras et al. utilized ML to classify patients with ICC into three distinctive groups using an
unsupervised hierarchical ML technique based on clinicopathologic characteristics: common type,
proliferative type, and inflammatory type . Notably, this classification was correlated with survival
[54]
outcomes with median OS values of 60.4 months for the common type, 27.2 months for the proliferative
type, and 13.3 months for the inflammatory type (P < 0.001).

Recently, other research teams have examined the integration of radiomics data into prediction models to
enhance discrimination ability. For instance, Chen et al. incorporated the 3D tumor region of interest (ROI)
derived from contrast-enhanced CT to predict very early recurrence of intrahepatic cholangiocarcinoma. In
this study, the K-means clustering algorithm was employed to identify novel radiomics-based subtypes of
intrahepatic cholangiocarcinoma . Notably, two distinct subtypes based on radiomics features were
[55]
identified, with subtype 2 tumors demonstrating a higher proportion of very early recurrence (VER) (47.6%)
versus subtype 1 lesions (25.5%).

Biliary tract cancer, particularly perihilar cholangiocarcinoma or central-type ICC, often necessitates
extensive surgery [56,57] . While major liver resection with vascular resection may lead to a high incidence of
morbidity and mortality, patients who require such radical resection might have limited survival benefits,
raising questions about the effectiveness of the procedure . To that end, the optimal allocation of
[58]
treatment in these patients has been a topic of debate. AI has emerged as a possible tool to address this
challenge. For example, Ratti et al. used an ML algorithm to identify patients most at risk for s “futile”
outcomes after surgery for hilar perihilar cholangiocarcinoma defined as severe complications with early
recurrence . Of note, independent predictors of futility included an American Society of Anesthesiology
[59]
(ASA) score ≥ 3, bilirubin at diagnosis ≥ 50 mmol/L, CA 19-9 ≥ 100 U/mL, preoperative cholangitis, portal
vein involvement, tumor diameter ≥ 3 cm, and left-sided liver resection. The ML-based scoring system
demonstrated good accuracy (AUC 0.755) in the validation cohort. The authors suggested that identifying
patients at high risk of “futility” using this AI approach may help guide the consideration of alternative
treatment options. In another study, Alaimo et al. utilized OPT analysis to define the optimal surgical
margin width based on individual clinicopathologic factors . The OPT categorized surgical patients into
[60]