The role of radiomics in hepato-bilio-pancreatic surgery: a literature review

Diagnosis

Transabdominal US, alone or in combination with serum markers, is the backbone for surveillance and early identification of HCC in high-risk subjects^[6,7], as it has a sensitivity of 94% for detecting HCC before it becomes clinically apparent; however, US has a lower sensitivity (63%) for detecting early-stage HCC^[8]. There are several reasons for this, including tumor location, size, and echogenicity, as well as patient-related factors that limit US exploration, such as poor cooperation, obesity, and marked steatosis, which are relevant given the increasing prevalence of non-alcoholic fatty liver disease (NAFLD). Enhancing the diagnostic capability of US for early identification of HCC would be highly beneficial from a clinical point of view. Radiomics has a potential role in this regard, although the evidence is very scarce: one single study by Yao et al. explored radiomic analysis in 177 subjects with a liver lesion who underwent B-mode US, shear wave elastography, and shear wave viscosity imaging; 2,560 features were extracted and five radiomic models were constructed to differentiate between benign and malignant lesions and to diagnose HCC^[9]. The areas under the curve (AUCs) were 0.94 for differentiation between benign and malignant lesions and 0.97 for classification of the malignant subtype. Although CT and MR are frequently performed after an initial insufficient US examination of the liver, their role in screening has not been investigated, as both are unsuitable due to radiation burden and high cost, respectively. Nine of the 54 studies included in the systematic review by Harding-Theobald et al. examined aspects of HCC diagnosis based on radiomics analysis, but most of them focused on the differentiation between hemangiomas from HCC, which is a simple task for trained Radiologists; moreover, these studies were of low methodological quality, with a RQS ranging from 5 to 10^[10]. Dankerl et al. trained a computer-aided diagnosis (CAD) system based on CT texture analysis in 372 patients with 2,325 liver lesions^[11]. The diagnostic performance of the CAD system for differentiation between benign and malignant lesions and between cysts, hemangiomas, and metastases was high, with AUCs of 94.5% for lesion type and 91.4% for lesion histology; overall, the CAD system performed better than three radiologists blinded to clinical information and with access only to CT images. Using the European Association for the Study of the Liver (EASL) guidelines, Mokrane et al. retrospectively examined 178 patients with cirrhosis with radiologically indeterminate, biopsy-proven liver nodules^[12]. 142 and 36 patients were randomly chosen for discovery and validation cohorts, respectively. Nodules were segmented on CT images, and 13,920 features were extracted. The radiomic signature was trained, calibrated, and validated using machine learning for differentiation between HCC and non-HCC nodules: a single radiomics feature had an AUC of 0.70 and 0.66 in the two cohorts. Laino et al. reviewed 11 studies on automatic classification of liver nodules using the Liver Imaging Reporting and Data System (LI-RADS) criteria: all studies demonstrated that radiomics performs well in the classification of liver nodules, sometimes better than human evaluation, reaching an AUC of 0.98 either on CT or MR^[13].

Prognostic stratification

Prognostic stratification is the most challenging and fascinating application of radiomics. This can be obtained either by correlating radiomics features to known adverse pathological features or by directly comparing radiomics features to overall survival (OS) and recurrence-free survival (RFS). Tumor grade has a role in determining the prognosis of HCC patients, as disease recurrence and metastasis are more likely to develop in high-grade HCCs^[14]. Mao et al.^[15] and Wu et al.^[16] reported high AUC values for combined clinical and radiomics models for HCC grading prediction (0.801 and 0.800, respectively). Microvascular invasion (MVI) has an impact on survival outcomes and disease recurrence. Studies conducted on the role of radiomics in predicting MVI reported AUC values ranging from 0.76 to 0.94^[10,17-20], with MR-based studies having the highest diagnostic value in predicting MVI. Another negative prognostic factor in HCC is vessels encapsulated tumor clusters (VETC), which are histologically defined as the presence of vessels marked with CD34 completely encapsulating neoplastic clusters. Yu et al. developed a predictive model based on contrast-enhanced MR images obtained after gadolinium ethoxy benzylic diethylenetriamine pentaacetic acid (Gd-EOB-DTPA) injection, with AUC of 0.972 for preoperative prediction of VETC^[21]. Finally, HCCs expressing biliary-specific markers such as CK19 have a higher rate of nodal metastasis and worse prognosis after surgery^[22]. Wang et al. developed a radiomics-based model derived from MR images with sensitivity and specificity values of 0.818 and 0.974 in the training cohort and 0.769 and 0.818 in the validation cohort^[23]. Disease recurrence after surgery negatively affects patients’ prognosis. According to 17 studies included in the meta-analysis by Harding-Theobald et al., radiomic models can predict early recurrence after surgery, with AUCs ranging from 0.71 to 0.86^[10]. The radiomics model developed by Ji et al. to predict early disease recurrence outperformed non-radiomics models and commonly used staging systems in terms of predictive performance (C-index ≥ 0.77) and prediction error (integrated Brier score ≤ 0.14)^[24]. The most advanced applications of radiomics incorporate immuno-molecular and genomic characteristics of HCC, which may aid in treatment stratification. Hectors et al. reported that qualitative and quantitative radiomics features were significantly correlated with several immunohistochemical markers such as CD3, CD68, CD31, PD-L1, PD1 and CTLA4^[25].

Response to treatment

Several radiological criteria have been proposed to assess treatment response in HCC patients, such as Response Evaluation Criteria In Solid Tumors (RECIST) and modified RECIST (mRECIST). Radiomics might enable a standardized, early and comprehensive evaluation of treatment response, with implications for patient management. Most studies were applied to patients who underwent loco-regional treatments for HCC^[10]. Kim et al. demonstrated that post-TACE OS could be predicted with a hazard ratio (HR) of 19.8 by combining clinical (Child-Pugh score, alphafetoprotein, and tumor size) and radiomics features (surface area-to-volume ratio, kurtosis, median, size zone variability)^[26]. Other studies reported similar results, indicating that radiomics features extracted from pre-treatment imaging were able to predict treatment response after TACE^[27]. Histogram analysis of apparent diffusion coefficient (ADC) maps seems to predict TACE response, as reported by Wu et al.^[28] and Shaghaghi et al.^[29]. Yang et al.^[30] developed a radiomics score composed of four features that independently affected recurrence after ablation at 1, 2, and 3 years with AUCs of 0.79/0.72, 0.72/0.61, and 0.71/0.64 in the train and validation groups, respectively.

Diagnosis

Liu et al. aimed to differentiate combined HCC-CC from iCC and HCC using MRI and CT radiomics features^[32]. MRI-based radiomics features were the best for differentiation of HCC-CC from non-HCC-CC (AUC = 0.77). Zhou et al. aimed to differentiate HCC-CC from mass-forming iCC by extracting radiomics features from contrast-enhanced MR images: a radiomics nomogram including alpha-fetoprotein, coexistent liver disease, and radiomics signature had an AUC value of 0.945 in the training cohort and 0.897 in the validation cohort^[33].

Staging

Even using high-end preoperative imaging, understaging is a problem, as a substantial proportion of patients with cholangiocarcinoma (CC) have occult metastases detected only on resection specimens or through early recurrence after resection. Moreover, positive resection margins are not uncommon. For these reasons, preoperative risk stratification should be as precise as possible to better stratify patients’ prognoses. Ji et al. developed a radiomics signature with significant association with nodal metastases in a cohort of 155 patients (test cohort = 103 patients; validation cohort = 52 patients); by adding CA 19.9 levels to the radiomics signature, the model had AUC of 0.846 and 0.892 in the two cohorts, respectively^[34]. Chu et al. included 203 iCC patients from two centers^[35]. Clinical and CT-derived radiomics features were selected to develop two models predictive of unbeneficial surgery (i.e., with macroscopic residual tumor or definitely unresectable). The radiomic model had higher AUC than the clinical model (0.804 vs. 0.590; P = 0.043) and reached 84.6% sensitivity and 77.1% specificity values. In the study by Qin et al., 274 patients who underwent contrast-enhanced CT and resection were divided into training (n = 167), internal validation (n = 70) and external validation (n = 37) cohorts^[36]. 18,120 radiomics features and 48 clinical and radiologic characteristics were analyzed. A model based on tumor differentiation, nodal metastasis, preoperative CA 19.9 level, tumor enhancement, the “shrink score” (i.e., features extracted from a ROI comprising 50% of the entire tumor area on arterial, portal and delayed contrast phases) had an AUC of 0.883 and performed better than clinical and radiomic models (AUCs 0.792-0.805); the model had an accuracy of 0.826, which was higher than AJCC 8th^[37], MSKCC^[38], and Gazzaniga^[39] staging systems (AUCs, 0.641, 0.617, and 0.581, respectively).

Prediction of prognosis

Some studies have been published concerning the possibility of predicting the prognosis of patients with CC. The largest series in this regard was reported by a multicenter, retrospective study by Park et al., including 345 patients with mass-forming iCC who underwent surgery^[40]. A clinical-radiologic model including infiltrative margins, multifocality, periductal infiltration, extrahepatic infiltration, and nodal metastases had similar performance in predicting RFS compared to the radiomics model (C-index, 0.65 vs. 0.68; P = 0.43). A clinical-radiological-radiomics model performed better than the clinical-radiologic model (C-index, 0.71; P = 0.01), with similar performance to commonly used postoperative prognostic systems to predict RFS (C-index, 0.71-0.73 vs. 0.70-0.73; P > 0.05) and OS (C-index, 0.68-0.71 vs. 0.64-0.74; P > 0.05). Zhang et al. developed an MRI-based texture signature predictive for the immunophenotyping and OS of patients with iCC: 78 patients were divided into two cohorts (inflamed iCC, n = 26; non-inflamed iCC, n = 52) based on CD8+ T cells density; arterial phase MR images were analyzed^[41]. A combination of three wavelets and one 3D feature were able to discriminate immunophenotyping (AUC = 0.919).

Treatment selection and response to treatment

The radiomic feature Wavelet-HLH_firstorder_Median was associated with OS, with a C-index of 0.70. CT texture analysis can quantify vascularization and homogeneity of iCC, providing useful information in identifying optimal candidates for trans-arterial radioembolization (TARE), as reported by Mosconi et al.: in this study, analysis was retrospectively performed in 55 pre-TARE CT scans; iCCs showing objective response after TARE had a higher mean histogram values (P < 0.001), GLCM homogeneity (P = 0.005) and GLCM correlation (P = 0.030), and lower kurtosis (P = 0.043), Grey-level co-occurrence matrix (GLCM) contrast (P = 0.004), and GLCM dissimilarity (P = 0.005) at the pre-TARE CT scan^[42].

Prediction of Kirsten Rat Sarcoma Virus gene (KRAS) status

Kirsten Rat Sarcoma Virus gene (KRAS) mutation is associated with worse prognosis; on the other hand, KRAS inhibitors are more and more commonly used for the treatment of metastatic colorectal cancer. Predicting KRAS status with non-invasive methods would be, therefore, clinically useful. A meta-analysis by Jia et al. reported on the prediction of the KRAS status in patients with colorectal LM, with pooled sensitivity, specificity and AUC values of 0.80/0.78, 0.80/0.84 and 0.87/0.86 in the training and validation cohorts, respectively^[45].

Response to treatment

Early and accurate response evaluation in patients with LM would be of importance given the availability of several different treatment modalities, and several studies^[47-50] reported a significant association between radiomics features and response to chemotherapy and targeted therapies.

Early detection of pancreatic cancer

Despite improvements in the multidisciplinary treatment approach to patients with pancreatic ductal adenocarcinoma (PDAC), the prognosis of this disease remains very poor, as PDAC is frequently diagnosed at an advanced stage. The goal of screening and surveillance programs is to detect and treat stage I PDAC and cancer precursor lesions as intraductal papillary mucinous neoplasms (IPMNs) with high-grade dysplasia^[54]. Even though encouraging results were described by two studies^[55,56], a meta-analysis by Chhoda et al. reported a significant proportion of PDACs diagnosed at a late stage during follow-up, which limits the survival benefit of surveillance^[57]. Although no studies have been conducted in this regard, AI and radiomics could theoretically play a role, as they can automatically differentiate between cancer and normal pancreas, as reported by Chu et al.: in this study, 427 features were extracted from CT images; overall, the accuracy of the binary random forest classification was 99.2%, with an AUC of 99.9%, a sensitivity of 100% and a specificity of 98.5%; a major limitation of this study is that the mean tumor size was 4.1 cm, which usually corresponds to a non resectable disease^[58]. Two studies by Qureshi et al.^[59] and Javed et al.^[60] identified several features that may be predictive of PDAC development extracted from pre-diagnostic CT scans in PDAC patients: the predictive model had an accuracy of 86%-89.3%, a sensitivity of 86% and a specificity of 93%; unfortunately, these results were based on a very limited study population and needed further validation. The most robust study on early automated detection of PDAC was published by Mukherjee et al.: they used a radiomics-based machine learning model to detect PDAC before the clinical diagnosis based on volumetric segmentation of the pancreas performed on pre-diagnostic CT scans in 155 PDAC patients and 265 normal subjects^[61]. A supporting vector machine model had high sensitivity (95.5%), specificity (90.3%), AUC (0.98), and accuracy (92.2%) for the classification of CT into pre-diagnostic versus normal. The paradigm for early detection of cancer in IPMN is based on “worrisome features” and “high-risk stigmata”. Accurate prediction of the malignant potential of IPMN is of great importance. Nevertheless, studies on AI/radiomic applications in identifying “malignant” IPMNs comprised very small study populations. The largest were proposed by Jeon et al.^[62] and Chakraborty et al.^[63], who reported that radiomics features improve the performance of MR and CT for predicting malignant IPMNs. Circulating micro RNAs (miRNAs) may be diagnostic biomarkers of incidentally detected IPMNs and predictors of their histological classification: a study by Permuth et al. suggested that the combination of radiomics features with the miRNA genomic classifier (MGC) had an AUC = 0.92 and sensitivity (83%), specificity (89%), PPV (88%), and NPV (85%) for prediction of IPMN malignancy^[64].

Differential diagnosis

Radiomics is a promising tool to improve the characterization of focal pancreatic lesions. Zhang et al. developed CT and radiomics nomograms for the differentiation between mass-forming pancreatitis (MFP) and PDAC in patients with chronic pancreatitis (CP)^[65]. 138 patients with histopathologically diagnosed MFCP or PDAC were retrospectively analyzed. Both models had good performance in differentiating between the two entities in the training (AUC = 0.87/0.91) and validation (AUC = 0.94) cohorts. A radiomics-based computer-aided diagnosis scheme proposed by Wei et al. had AUC = 0.767, sensitivity = 0.686, and specificity = 0.709 for preoperative diagnosis of cystic neoplasms^[66].

Prediction of prognosis

Tumor grade is a major prognostic factor for pancreatic neoplasms. Chang et al.^[67] and Tikhonova et al.^[68] reported significant differences in radiomics signatures between PDAC of different grades, with AUC ranging from 0.66 to 0.77. Tumor grade is even more important from a prognostic point of view for pNENs. Gu et al. included 138 patients with pathologically confirmed pNENs (training cohort, 104 patients; validation cohort, 34 patients)^[69]. A nomogram integrating tumor margin status and a radiomic signature derived from CT images showed strong discrimination with AUCs of 0.974 and 0.902 in the training and validation cohort, respectively.

Several studies conducted a radiomic analysis to predict features with a negative prognostic role of pNENs. In particular, a study^[70] retrospectively analyzed the MR histogram features of 42 patients with pNEN: ADCentropy was higher in G2-3 tumors with AUC = 0.757, sensitivity and specificity of 83.3% and 61.1%, while kurtosis was higher in pNENs with vascular infiltration, lymph node and liver metastases (P = 0.008, 0.021 and 0.008; AUC = 0.820, 0.709 and 0.820). Mori et al. evaluated radiomics features extracted from unenhanced CT images and reported AUCs of 0.81/0.81 for the radiomic and clinic-radiological model for metastases and 0.67/0.72 and 0.68/0.70 for tumor grade^[71]. Transcriptional classifiers are key prognostic factors of PDAC. Salinas-Miranda et al.^[72] developed a radiomics score that was significantly associated (coefficient = 0.31) with the four PDAC subtypes (squamous, pancreatic progenitor, immunogenic, and aberrantly differentiated endocrine exocrine) classified by Bailey et al.^[73] through gene expression analysis. Preoperative staging of pancreatic neoplasms includes the evaluation of vascular involvement, node and distant metastases. Rigiroli et al. developed a 5-feature model with AUC = 0.71 for preoperative assessment of superior mesenteric artery involvement by PDAC^[74]. Bian et al. developed a radiomic score that was linked to the possibility of nodal metastasis^[75]. Finally, a study that comprised 220 patients^[76] developed a logistic regression model composed of eight variables that had sensitivity, specificity, PPV, NPV and AUC of 58.6%, 91.3%, 75.9%, 82.5%, and 0.850 for prediction of simultaneous liver metastases in PDAC patients. Early recurrence after surgery is not uncommon in PDAC patients. Although the reasons for this have not been defined with certainty, the availability of a non-invasive biomarker would be clinically useful in patients with resectable PDAC to avoid unbeneficial surgery. Tang et al. analyzed 303 PDAC patients; the AUC values for prediction of early recurrence of radiomics signatures were 0.80, 0.81, and 0.78 in the training, inner validation, and outer validation cohorts, respectively; the AUC values for early recurrence were 0.87, 0.88, and 0.85 in the training, internal validation, and external validation cohorts^[77]. A study aimed to correlate conventional and radiomics MR features with the risk and the time to metastases after surgical resection in patients with resected PDAC^[78]; 120 patients were included. ADC skewness had a significant impact on the risk of metastases, with HR = 5.22 (P < 0.001): the time to metastases was significantly shorter (11.7 vs. 30.8 months, P < 0.001) in patients with an ADC skewness value greater than 0.23. Two studies^[79,80] reported the association between radiomics features and adverse pathological features, including margin and nodal status, tumor grade, lymphovascular and perineural invasion. Several studies reported the correlation between radiomics features and clinical outcomes in terms of RFS and OS^[81-84]. Few data are available on the potential usefulness of radiomics to assess treatment response. Borhani et al.^[85] and Kim et al.^[86] reported that textural features extracted from CT could be used as biomarkers predictive of the histologic response, the DFS and the OS after neoadjuvant chemotherapy. Cozzi et al.^[87] developed a clinical-radiomic signature which was significantly associated with OS in PDAC patients treated with stereotactic body radiation therapy (SBRT) in training and validation cohorts (P = 0.01 and 0.05; concordance index 0.73 and 0.75, respectively).