Radiomics might offer an advantage over other techniques in the early detection of PDAC. This is because the subtle difference of the texture patterns between early cancer and normal pancreas might be discernable using radiomic features prior to visual detection.

Chu et al[18] used 3D CT radiomic features to differentiate PDAC and normal pancreas by manually segmented features of the pancreas. The dataset included 190 patients with PDAC and 190 healthy controls, and was divided into 255 training and 125 validation cases. A total of 478 features was extracted, and 40 features were selected for analysis by a random forest (RF) classifier. The overall accuracy was 99.2%, and the area under the curve (AUC) was 99.9%. The results were encouraging for using radiomics in the early detection of PDAC, but a limitation of this study was that the manual segmentation of pancreas boundaries was a labor-intensive work and required expert knowledge of radiologists.

To overcome this limitation, Liu et al[19] used CNN to distinguish 370 patients with pancreatic cancer and 320 normal controls. CT images were preprocessed into patches to classify as cancerous or non-cancerous. In local test sets, CNN-based analysis had an accuracy of 0.986–0.989 and AUC of 0.997–0.999. In the test set (281 pancreatic cancers and 82 controls) of a different country, the accuracy was 0.832 and AUC was 0.920. The sensitivity for tumors smaller than 2 cm was 92.1% in the local test sets and 63.1% in the other country test set. When compared with radiologists’ interpretation, CNN-based analysis achieved higher sensitivity than radiologists. Therefore, this method could be incorporated into the development of computer-aided detection software for pancreatic cancer detection. In clinical practice, other benign lesions, such as MFP or AIP, might mimic PDAC. Whether CNNs can distinguish between PDAC and other pancreatic pathologies, such as pancreatitis and other pancreatic tumors, must also be further studied. Besides, about 11%–27% of pancreatic cancer is enhancing the pancreatic parenchyma and not visible on contrast-enhanced CT[20]. It is interesting to see whether radiomics can detect this particular type of PDAC.

Accurate identification of the extent of lymph node (LN) metastasis is critical for the determination of surgical methods in resectable PDAC.

Li et al[21] developed a model integrating clinical data and imaging features extracted from venous phase CT to predict LN metastasis. Their study included 159 patients with PDAC (118 in the primary cohort and 41 in the validation cohort). A total of 2041 radiomics features were extracted, and 15 features were selected for constructing the radiomics signature in the primary cohort. A combined prediction model was built by integrating the radiomics signature and clinical characteristics selected by using multivariable logistic regression. The combined prediction model reached a better discrimination power than the clinical prediction model, with an AUC of 0.944 vs 0.666 in the primary cohort, and 0.912 vs 0.713 in the validation cohort.

Bian et al[22] used arterial phase CT images to predict LN metastasis in 225 patients. A total of 1029 radiomics features of the arterial phase were extracted and then reduced using the least absolute shrinkage and selection operator logistic regression (LASSO) algorithm. Multivariate logistic regression models were used to analyze the association. The radiomics score (rad-score), which consisted of 12 selected features, was significantly associated with LN status, both in univariate and multivariate analyses. Higher arterial rad-score was also associated with LN metastasis. In the future, it is necessary to establish a one-to-one correlation between the imaging findings and the pathological evidence of LN metastasis.

In a pathological examination after pancreaticoduodenectomy (PD), a resection margin without cancer cells in 1 mm is considered as R0; a resection margin with cancer cells in 1 mm is considered as R1. The preoperative identification of R0 and R1 is a determining factor for surgical decisions and prognosis[23,24].

Hui et al[25] retrospectively analyzed CT images of 86 patients (34 cases of R0 and 52 cases of R1) with pancreatic head PDAC and that underwent PD. The radiomics features were reduced using principal component analysis. The support vector machine (SVM) with a linear kernel was used to classify the resection margins with leave-one-out cross-validation. The results achieved an AUC of 0.8614 and an accuracy of 84.88%. Two features of the run-length matrix, which are derived from diagonal sub-bands in wavelet decomposition, showed significant differences between R0 and R1.

Similarly, Yun et al[26] used a portal rad-score to predict pathologic superior mesenteric vein (SMV) resection margin in 181 patients. For each patient, 1029 radiomics features of the portal phase were extracted, which were reduced using the LASSO logistic regression algorithm. The rad-score was significantly associated with the SMV resection margin status. The portal rad-score had an accuracy of 71.3% and AUC of 0.750. Although radiomics seem promising in predicting SMV section margin, assessment of all pancreatic resection margins is needed to predict patients’ outcomes. Furthermore, the radiomic features of mesopancreas (located between the superior mesenteric artery and the uncinate process) are more likely to predict the status of the section margin than those of a primary tumor, because it is regarded as the primary site of cancer cell infiltration[27].

Zhang et al[28] used radiomic features extracted from the portal venous phase CT for the preoperative prediction of postoperative pancreatic fistula (POPF) in 117 patients receiving PD. The rad-score was constructed by LASSO, and its performance was compared with standard pancreatic Fistula Risk Score. Their rad-score could predict POPF with an AUC of 0.8248 in the training cohort (80 patients) and of 0.7609 in the validation cohort (39 patients). In addition, the AUC of the rad-score was statistically higher than the Fistula Risk Score for predicting POPF in both cohorts.

Many researchers have utilized radiomic features derived from pretreatment CT to identify imaging phenotypes that might predict the treatment response in patients with PDAC.

Chen et al[29] assessed the response of pancreatic head cancer during chemoradiation therapy in 20 patients. They found that significant changes in CT radiomic features were observed during therapy based on quantitative analysis of daily CT. In cases of good response, patients tend to have large reductions in mean histograms of CT number and skewness, and large increases in standard deviation and kurtosis. Thus, a high reduction of these features might suggest early treatment response and could be used to identify patients that need therapeutic intensification.

Borazanci et al[30] used texture analysis to predict treatment response to poly adenosine diphosphate-ribose polymerase (PARP) inhibitors. In 13 patients with PDAC who have deoxyribonucleic acid damage repair deficiency mutations, exploratory analysis of index lesions revealed correlations between lesion texture features with overall survival (OS), and also with time on PARP inhibitors.

Yue et al[31] stratified patients into low and high-risk groups using pre- and post-radiotherapy 18F-FDG-PET/CT images from 26 patients. A total of 48 texture and clinical variables were identified, and the prognostic heterogeneity features were selected using LASSO/elastic net regression and multivariate Cox analysis. After radiotherapy, the metabolic activity in the primary tumor was suppressed, and underlying tissue heterogeneity was reduced. The authors identified five significant variables: Age, node stage, variations of homogeneity, variance, and cluster tendency. These patients could be stratified into two risk groups: A low-risk group (n = 11) with a longer mean OS and higher texture variation (> 30%), and a high-risk group (n = 15) with a shorter mean OS and lower texture variation (< 15%). The authors concluded that locoregional metabolic texture response might predict clinical outcomes following radiotherapy.

Recent studies have suggested that radiomic features extracted from CT and PET were predictive of the survival outcome of PDAC patients.

Xie et al[32] developed a CT-based radiomics nomogram for survival prediction in patients with resected PDAC in 220 patients (training = 147; validation = 73). A total of 300 radiomic features were extracted, followed by LASSO with multivariate regression analysis. The rad-score was significantly associated with disease-free survival (DFS) and OS. Radiomics nomogram could better predict survival than the clinical model, and the TNM staging system could. However, there was no association between the rad-score and recurrence patterns.

Cozzi et al[33] used CT radiomics signature to predict clinical outcomes after stereotactic body radiation therapy in 100 patients (training = 60; validation = 40) and found a clinical-radiomics signature was associated with OS and local control.

The value of texture features to predict prognosis and help clinical management in PDAC patients has been evaluated in several studies. In patients undergoing surgical resection, Kim et al[34] found that high grey-level non-uniformity values suggested shorter recurrence-free survival in 116 patients, suggesting that high tumor heterogeneity was a poor prognostic indicator. However, Yun et al[35] found that lower average values with homogeneous features (lower standard deviation and contrast and higher correlation) were significantly associated with poorer DFS in 18 patients. They conjectured that homogeneous texture features could represent more aggressive tumor nature, resulting from higher cellular density or dense desmoplasia. Besides, Eilaghi et al[36] found that high tumor dissimilarity (high heterogeneity) and low inverse difference normalized (low heterogeneity) were associated with better OS in 30 patients. Therefore, the results of correlations between tumor heterogeneity with surgical outcome were contradictory and need further investigation.

In patients with unresectable PDAC treated with chemotherapy, Cheng et al[37] found pretreatment CT texture analysis was associated with PFS and OS in 41 patients. Besides, a combination of pretreatment standard deviation (spatial scaling factor = 3) with tumor size in the survival model performed better than the standard deviation alone. Similarly, Sandrasegaran et al[38] found that texture features of the mean value of positive pixels and kurtosis at medium spatial filters had a significant correlation with OS in 60 patients.

Hyun et al[39] evaluated intratumoral heterogeneity measured by 18F-FDG PET texture analysis in 137 patients. The best imaging biomarker for OS prediction was first-order entropy (AUC = 0.720), followed by total lesion glycolysis (AUC = 0.697), metabolic tumor volume (AUC = 0.692), and maximum standard uptake value (AUC = 0.625). Multivariable Cox analysis demonstrated that higher entropy was independently associated with worse survival. Thus, first-order entropy is a better quantitative imaging biomarker of prognosis than conventional PET parameters.

AIP has similar clinical and radiological presentations to PDAC, but the treatments of these two entities are different. Patients with AIP might be treated with oral corticosteroids, but patients with PDAC need surgical resection and chemotherapy. Thus, the differentiation of these two entities is imperative to avoid unnecessary surgical resections in patients with AIP or delayed treatment in patients with PDAC.

Park et al[54] used CT-based machine learning of radiomic features to distinguish AIP from PDAC. Eighty-nine patients with AIP and 93 patients with PDAC were retrospectively included. Four-hundred-thirty-one radiomic features were extracted, and a RF method was used to discriminate AIP from PDAC. The radiomic features help differentiate AIP from PDAC with a sensitivity of 89.7%, specificity of 100%, accuracy of 95.2%, and AUC of 0.975.

Zhang et al[55] used 18F FDG PET/CT to distinguish AIP from PDAC in 111 patients (AIP = 45, PDAC = 66). They extracted 251 features from 2D and 3D images and recombined these features into five feature sets according to their modalities and dimensions. Four machine learning classifiers were evaluated. CT features and 3D features performed better than PET features and 2D features, respectively. Multidomain features were superior to single domain features. In addition, the combination of the SVM-recursive feature elimination feature selection strategy and linear SVM classifier had the best performance (AUC = 0.93, accuracy= 0.85). The radiomics model was significantly superior to both human doctors and clinical factors-based prediction models.

The results of these studies are encouraging. For future work, combined features extracted from CNNs and more clinical factors to differentiate these two diseases would be an interesting direction to pursue.

Ren et al[56] used arterial and portal phase CT texture analysis to differentiate 30 patients with MFP and 79 patients with PDAC. Arterial CT attenuation, arterial, and portal enhancement ratios of MFP were higher than PDAC. Arterial CT attenuation and pancreatic duct penetrating sign were independent predictors in multivariate analysis. AUC of imaging feature-based, texture feature-based in arterial and portal phases, and the combined models were 0.84, 0.96, 0.93, and 0.98, respectively. Thus, CT texture analysis holds great potential to differentiate MFP from PDAC.

Mashayekhi et al[57] used CT radiomics to differentiate 56 patients with recurrent acute pancreatitis (n = 20), functional abdominal pain (n = 19), or chronic pancreatitis (n = 17). In 54 radiomic features extracted by one-vs-one Isomap SVM classifier, 11 radiomic features were significantly different between the patient groups with an overall accuracy of 82.1%.

Yang et al[58] used CT textural features in the differential diagnosis of pancreatic serous cystadenomas (n = 53) and mucinous cystadenomas (n = 25). Textural parameters were analyzed using RF and LASSO methods. Patients were divided into training and validation sets with a ratio of 4:1. Radiomic features were able to separate serous from mucinous cystadenomas in both the training group (slice thickness of 5 mm, AUC 0.72, accuracy 0.86) and the validation group (AUC 0.75, accuracy 0.83). These results might provide a noninvasive approach to determine whether surgery or imaging follow up is suitable for these patients.