The extreme aggressiveness of pancreatic cancer greatly dampens survival probability when the cancer is diagnosed in late stages, with a five-year survival rate of 3% for metastatic disease ^[1]. Unfortunately, pancreatic cancer is usually not detected until the late stages, with metastatic pancreatic cancer counting for about 52% of patients ^[2]. In contrast, while early-stage resectable pancreatic cancer has a much better five-year survival rate of 39%, only 11% of patients are detected at this stage ^[3]. Early detection of pancreatic cancer is highly valuable in improving pancreatic cancer survival; however, early detection is challenging, as there are no validated screening tests available for pancreatic cancer. Current efforts focus on risk stratification based on intraductal papillary mucinous neoplasms (IPMNs) and pancreatic intraepithelial neoplasia (PanIN) as well as familial risk factors ^[4]. Pancreatic lesions that are unlikely to progress to cancer may not be good candidates for surgical resection, as the operation is highly risky, while precancerous lesions may, as their prognosis is worse. Because of the potential toxicity and mortality associated with invasive biopsy of the pancreas, screening and early detection relies heavily on medical imaging. However, the small size of the early lesions/precursors and complex radiological appearances of these lesions and their background structures substantially challenge conventional radiology in providing reliable image-based early detection and diagnosis of precursor lesions. This offers a window of opportunity for novel quantitative imaging approaches. Both radiomics and deep learning methods have been applied in these types of applications.

A series of papers were identified applying radiomics in pancreatic precancerous lesion diagnosis. As discussed in the introduction, manual segmentation was used in most of these studies. For automatic and semi-automatic segmentation the software or algorithm used to automate the segmentation. Readers are referred to the original publications for additional details on segmentation. Standard-of-care imaging modalities (primarily CT) were used in these studies, though other modalities such as PET were used as well. For example, several radiomic models were developed to diagnose and evaluate the malignancy of cysts, showing improved accuracy compared to conventional radiology ^{[5][6][7][8][9][10][11][12][13][14][15][16][17][18]}. Pancreatic cysts have a fairly common occurrence relative to pancreatic cancer in adults; prevalence increases with age, and can show up as incidental image findings or on screening images. Because these cystic lesions are correlated with a wide range of histologic differentiation and malignant risk, image diagnosis distinguishing among them represents a crucial challenge. The conventional radiological diagnosis of pancreatic cysts is only accurate 60–70% of the time ^[19]. Using quantitative approaches, studies have been able to create a radiomics-based model in order to differentiate cyst types and propose risk stratification useful in determining treatment. High-risk lesions are recommended as candidates for surgical resection, while low-risk lesions are recommended as candidates for less aggressive management. For example, Wei et al. retrospectively studied CT-based radiomics on 260 patients with pancreatic cystic neoplasm and who underwent a pancreatic resection for it ^[8]. Using the pathology-established diagnosis of these 260 lesions, the researchers grouped them into benign serous cystic neoplasms (SCNs) and malignant non-SCNs, the latter consisting of IMPNs, mucinous cystic neoplasms, and solid pseudopapillary neoplasms. With cross-validation in the training cohort of 200 patients, their radiomic model achieved an area under the receiver operating characteristic curve (AUC) of 0.84 on the independent validation of 60 patients. In contrast, only 30% (31 of 102) of the SCNs were correctly diagnosed pre-surgery with conventional radiology; thus, radiomics clearly shows potential as a computer-aided diagnosis (CAD) tool for improving the efficiency and accuracy of pancreatic cyst diagnosis. In another study using 53 patients with surgically resected IPMNs, Hanania et al. showed that a radiomic model based on texture features could differentiate low-grade versus high-grade IPMNs for the risk stratification necessary in clinical workflows and treatment decisions, with a high AUC of 0.96 in cross-validation compared with a false positive rate of 36% using the clinical Fukuoka criteria ^[5]. Huang et al. were able to use a radiomic model to predict invasive behavior in pancreatic solid pseudopapillary neoplasms ^[20]. Song et al. used a radiomic model to predict the recurrence risk of pancreatic neuroendocrine neoplasms ^[21]. Watson et al. applied a deep learning model to predict the malignancy of pancreatic cystic neoplasms ^[22]. Surgery is the only treatment for these neoplasms, and using radiomic and deep learning models can help to predict the prognosis and therefore led to a more accurate clinical decision regarding surgical resection. The current progress on texture analysis of pancreatic lesions for differential diagnosis has been reported in a recent review by Awe et al. ^[23].

Deep learning models are able to differentiate and risk stratify precancerous pancreatic lesions as well. Similar to radiomics studies, most of these studies were based on CT, though other modalities such as MRI and EUS were studied as well. Several studies adopted CNNs to predict lesion or cyst diagnosis and malignancy, such as lesion type or grade ^{[24][25][26][27][28][29][30]}. For example, using the EUS images of 206 patients with IPMNs that were later surgically resected, Kuwahara et al. developed a CNN model that achieved 94% accuracy for IPMN malignancy diagnosis, compared with 56% accuracy of human diagnosis ^[30]. In another study, Corral et al. used CNN to classify IPMN based on MRI images, and achieved a comparable AUC of 0.78, compared with 0.76 using the American Gastroenterology Association guidelines and 0.77 using the Fukuoka criteria ^[29]. In addition to these studies that used deep learning models alone, other studies combined deep learning with radiomics by adding radiomic features to the input channels of the deep learning algorithm, and others created fusion models (ensemble models) to integrate radiomics-based and deep learning-based predictions. Dmitriev et al. presented such a study ^[28]; using CT images of 134 patients with pancreatic cysts consisting of four histopathological types, they trained a radiomics model, a CNN model, and an ensemble/fusion model to classify the cyst lesion types ^[28]. On cross-validation, the radiomics model and the CNN model achieved an overall accuracy of 79.8% and 77.6%, respectively, while the ensemble/fusion model reached 83.6% ^[28]. Fusion models outperform radiomic models and deep learning models in these early lesion classification and malignancy diagnosis applications.

Because reliable pancreatic cancer detection and diagnosis is unattainable based simply on symptoms and signs, medical imaging plays an essential role. A variety of imaging modalities can be used, including transabdominal US, CT, ERCP, MRCP, etc. CT is the most commonly used imaging modality among these, with a reported sensitivity of detection ranging from mid~70% to high~90% ^[31]. However, the accuracy of detection and diagnosis is highly dependent on the radiologist’s experience; misdiagnosis and missed diagnosis are not uncommon. Therefore, radiomics and deep learning have been explored to aid the clinical task of image-based pancreatic cancer detection and diagnosis. 

For detection, several works have shown the utility of radiomic models in differentiating pancreatic cancer tissue and healthy tissue ^[32][33]. Chu et al. trained a whole-pancreas ROI-based radiomic model on 225 training cases and validated the model on 125 validation cases ^[33]; the resulting model consisted of 40 radiomic features and achieved a very high AUC of 99.9% ^[33]. Chen et al. applied radiomic features and machine learning to investigate the utility of radiomics modeling in detecting pancreatic cancer ^[32]. Based on contrast-enhanced CT, they observed that pancreatic cancer tends to be hypodense and more heterogeneous compared with normal pancreas, as reflected by the relevant radiomic feature values. Quantitatively, their radiomics model trained on >1000 subjects achieved AUCs of 0.98 and 0.91 on local and external test datasets.

Deep learning models have proven useful for detecting pancreatic cancer. Zhang et al. used a novel deep learning framework consisting of Augmented Feature Pyramid networks, Self-adaptive Feature Fusion, and a Dependencies Computation Module to detect pancreatic cancer tumors, which resulted in an AUC of 0.95 on internal testing ^[34]. In a large-cohort study, Liu et al. applied CNN-based modeling on contrast-enhanced CTs of ~700 subjects (~1:1 with pancreatic cancer vs. healthy pancreas controls, divided 4:1 into training and validation sets) ^[35]. The model was further tested on a ~200-subject independent local cohort and a ~350-subject independent international cohort, achieving an AUC of 1.00 and 0.92, respectively ^[35]. For their local validation and testing datasets, the performance of the CNN model was compared against that of human radiologists and showed significantly higher sensitivity ^[35]. In another study, Chu et al. reported their initial experience of training deep learning networks to detect pancreatic adenocarcinoma ^[36]. They took a two-step approach to their curated large CT cohort of pancreatic cancer patients and controls with a healthy pancreas, using supervised learning to first train deep learning models to automatically segment all abdominal organs, and then to detect pancreatic cancer. Their algorithms achieved segmentation performances superior to published state-of-the-art algorithms, and showed 94.1% sensitivity and 98.5% specificity for pancreatic cancer detection in preliminary testing.

As pancreatic cancer is highly aggressive, correctly diagnosing it from benign or less-aggressive lesions could reduce unnecessary surgical resections that potentially lead to patient morbidity. Image-based grading and histopathology prediction can aid better treatment stratification. Reinert et al. identified CT-based textual features that can differentiate between pancreatic ductal adenocarcinoma and pancreatic neuroendocrine tumors as well as between low-grade and high-grade pancreatic neuroendocrine tumors ^[37]. Gu et al. were able to differentiate pancreatic ductal adenocarcinoma and neuroendocrine tumors from solid pseudopapillary neoplasm using MRI radiomic features ^[38]. Zhao et al. and Benedetti et al. used CT-based radiomic features to discriminate pancreatic neuroendocrine tumor grades, and Bendetti et al. predicted lymph node invasion status ^[39][40]. Similarly, using texture features on CT, Canellas et al. were able to differentiate grade 1 from grade 2 and 3 pancreatic neuroendocrine tumors ^[41]. In a large-cohort study, Chang et al. used contrast-enhanced CTs of ~300 patients to train and validate radiomics models to differentiate high-grade versus low-grade pancreatic ductal adenocarcinoma, achieving an AUC of 0.91 on the internal validation set and 0.77 on a 100-patient external testing cohort ^[42]. Other imaging modalities have been explored as well; in an early study, Zhu et al. used a support vector machine model on texture features from EUS images to differentiate pancreatic cancer from chronic pancreatitis ^[43]. With a total of ~400 patients, >90% accuracy was achieved in cross-validation. Based on various MRI sequences, Deng et al. tested radiomics models and compared them against a clinical model to differentiate pancreatic cancer from mass-forming chronic pancreatitis ^[44]. The radiomic models developed on a training cohort achieved performances much superior to the clinical model on a validation cohort from a different institution (AUCs of 0.88–0.96 vs. 0.65). Bevilacqua et al. used [68Ga]Ga-DOTANOC PET/CT radiomic features to detect grade 1 and 2 pancreatic neuroendocrine tumors ^[45].

Deep learning models have been successfully applied in pancreatic cancer diagnosis. CNN-based models have been frequently used. Si et al. applied ResNet models to diagnose different pancreatic lesions and achieved an average accuracy of 82.7% for all tumor types ^[46], and generated saliency maps to highlight the image regions relevant to the decision. Saftoiu et al. used an extended neural network analysis on EUS elastography for differential diagnosis of pancreatic cancer and chronic pancreatitis, and achieved an average testing performance of 95% on cross-validation ^[47]. Ziegelmayer et al. applied CNN modeling on CT images to discriminate between pancreatic cancer and autoimmune pancreatitis, and compared it with radiomics modeling ^[48]. On cross-validation, they achieved an average AUC of 0.90 with the deep learning model, outperforming their radiomic model, which had an AUC of 0.80.

Correctly detecting and diagnosing pancreatic cancer is crucial for managing the disease and selecting optimal treatment, and is ultimately crucial for patient outcomes. Accurate diagnosis of non-cancer versus cancer could reduce unnecessary surgical resections, which can lead to patient morbidity. Overall, both radiomic and deep learning models show great promise in these clinical tasks and could be implemented into computer-aided diagnosis systems for pancreatic cancer. These models could reduce the time and manual effort involved in these clinical tasks, reduce invasive biopsy procedures, and potentially offer more accurate diagnoses in order to improve treatment planning and improve patient outcomes.

Significant radiomic and deep learning features in pretreatment images have been used to predict treatment efficacy for pancreatic cancer. Survival can be predicted by both radiomic and deep learning models using pretreatment images in order to determine the level of treatment needed for the best chance of patient survival. Similarly, predictions on recurrence, metastasis, and surgical margins can be used to strategize regarding the optimal treatment for an individual patient.

Radiomic models have been used to predict progression-free survival, relapse-free survival, and overall survival for pancreatic cancer ^{[49][50][51][52][53][54][55][56][57][58][59][60][61][62][63]}, and deep learning models have been used to model survival as well ^{[64][65][66][67][68]}. Mapelli et al. used a radiomic model to predict the aggressiveness of pancreatic neuroendocrine neoplasms ^[69]. Gao et al. used a deep learning CNN model to predict grades of PNET, with an average accuracy of 85.1% ± 0.4% and an average accuracy on the external validation set ranging between 79.1 and 82.4% ^[68]. Klimov et al. were able to predict metastasis risk in PNETs using a deep learning approach that had high ability to differentiate tumors, with an accuracy of >95% ^[70]. The models of both Gao et al. and Klimov et al. show promise in accurately assigning treatment plans to PNET patients while correctly predicting prognosis. Tang et al. were able to predict early recurrence in resectable pancreatic cancer using a radiomic nomogram ^[71]. Patients with a high risk of recurrence may be treated more aggressively or with other treatment modalities using the risk stratification proposed by the model developed by Tang et al. Bian et al. predicted the risk of lymph node metastasis in pancreatic cancer patients using a radiomic model ^[72]. Liu et al. were able to predict lymph node metastasis with their radiomic model as well ^[73]. If a pancreatic cancer patient already has lymph node metastases, a radical pancreatic operation may be futile and not worth the risks. The models developed by Bian et al. and Liu et al. could help to relieve patients from unnecessary operations.

In a different study, Bian et al. created a radiomic model to predict superior mesenteric vein resection margin (R1/2 vs. R0) in patients with pancreatic head cancer; their model was able to predict patient prognosis, as R1/2 resection is associated with poor overall survival relative to R0 resection ^[74]. Hui et al. were able to predict resection margin for pancreatic head cancer using a radiomic model as well ^[75]. Zhang et al. created a radiomics model that could predict postoperative pancreatic fistula in patients undergoing pancreaticoduodenectomy; their model could help with decisionmaking regarding risks of a pancreaticoduodenectomy ^[76]. Postoperative pancreatic fistula is one of the more harmful consequences of a pancreatic resection or pancreaticoduodenectomy, and the use of any of the aforementioned models could help to predict patient prognosis and assist with clinical decision making.

Li et al. created a radiomic model that could predict CD8+ tumor-infiltrating lymphocyte expression levels in patients with pancreatic ductal adenocarcinoma ^[76]. Patients with high levels of CD8+ tumor-infiltrating lymphocyte expression are possibly able to undergo immunotherapy targeting immune checkpoint inhibitors to improve their prognosis ^[77]. Bian et al. were able to predict tumor-infiltrating lymphocyte expression in their radiomic and deep learning models as well ^[78].

Apart from diagnosis and differentiation of pancreatic diseases, radiomics can play a helpful role in predicting optimal therapy paths. Several studies in radiomics have predicted treatment response to pancreatic cancer ^{[50][79][80][81][82][83][84][85][86]}. Parr et al. applied radiomic models to pretreatment CT images in order to predict overall survival and locoregional recurrence of pancreatic cancer after stereotactic body radiation therapy ^[50]. Their radiomic model and the model combining radiomic and clinical features outperformed the pure clinical model in these predictions, with an average concordance index of 0.66 and 0.68 versus 0.54 for survival and an average AUC of 0.78 and 0.77 vs. 0.66 for recurrence on validations ^[50]. Cozzi et al., using a hybrid clinical–radiomics model, were able to differentiate high and low risk in terms of overall survival for patients treated with stereotactic body radiation therapy, with an AUC of 0.81 ^[80]. Those patients with a low overall survival prediction may need more aggressive treatment, and with this model high-risk and low-risk groups can be more accurately identified. Watson et al. applied deep learning CNN to predict pathologic tumor response to neoadjuvant therapy in pancreatic cancer, with an AUC of 0.738 in predicting response to chemotherapy and an accuracy of 78.3% in predicting response to resectability ^[87].

Another direction of radiomics, as in studies on other cancers, is delta-radiomics. Delta-radiomics assesses the temporal change or kinetics of radiomic signatures and explores its value in evaluating tumor progression or predicting long-term patient outcomes. The examined temporal window using delta-radiomics can be relatively short, such as when using daily imaging during a radiation therapy treatment course, or more extended, as when using periodical imaging from diagnosis to treatment and follow-up. Delta-radiomics has been explored for pancreatic cancer thanks to its value in prognosis prediction and treatment stratification. Using daily imaging during a radiation therapy treatment course, Chen et al. showed that patients with good pathological pancreatic tumor response tended to have large changes in certain radiomic features of the tumor compared to those with poor tumor response; radiation-induced delta-radiomics may potentially be used for early assessment of treatment response during radiation delivery ^[88].

Aside from the above clinical applications, radiomics has been used in radiogenomics. Radiogenomics is an offshoot branch of radiomics that applies radiomic workflow and imaging features coupled with genomic profiles ^[89] to assess the association between imaging phenotypes and the underlying tumor biology. These radiomic signatures associated with underlying patterns of gene expression can then be used to predict prognosis and optimal treatment. While radiogenomics has been more widely studied for other types of cancer, pancreatic cancer radiogenomic studies remain sparse. On the other hand, the genomic landscape of pancreatic cancer is diverse and many mutations have been detected, creating many potential opportunities for radiogenomic analysis ^[90]. A few radiogenomic studies have been conducted ^{[91][92][90][93][94]}. Katabathina et al. suggest that the varied biological tumor features related to the different mutations in panNENs may result in morphological changes that are appreciable with imaging ^[95]. McGovern et al. identified CT-based radiomic features that are significantly associated with the alternative lengthening of the telomere phenotype of pancreatic neuroendocrine tumors ^[90]. Attiyeh et al. demonstrated that radiomic signatures of preoperative CT can predict the mutation status of certain pancreatic cancer driver genes, such as SMAD4 ^[90]. The use of radiogenomics can increase personalized medicine within pancreatic cancer patients, leading to better outcomes.

This entry is adapted from 10.3390/cancers14071654