Page 154 - Read Online

Grewal et al. Art Int Surg 2023;3:217-32 https://dx.doi.org/10.20517/ais.2023.28 Page 221

features have demonstrated an accuracy of 85%-95.2% in differentiating AIP from PC [30-34] . Similar studies
have been performed to differentiate PNETs from PDACs and SPT, achieving AUCs ranging from 0.86-
0.99 [35-38] . Of note, these studies included atypical hypovascular NF-PNETs, which more closely mimic
PDAC than typical PNETs, and radiomics analysis outperformed clinic-radiological factors [36,37,39,40] .
Radiomics combined with machine learning methods would potentially result in the development of tools
that will allow for radiomics-based screening for asymptomatic pancreatic cancer. Though the sensitivity of
CT in the detection of PDAC ranges from 76%-96%, early CT findings of PDAC, such as tumoral
heterogeneity and loss of fatty marbling, can be particularly subtle and may be missed even by experienced
radiologists . Radiomics poses an avenue for quantitative analysis and detection of these changes.
[28]
Radiomics analysis may also autonomously run in the background of scans and automate the process of
[41]
screening . This can effectively enable every abdominal CT scan, regardless of indication, to be used to
screen for PDAC. The improved quantitative analysis of images combined with an increase in the sheer
volume of scans being screened through this approach makes radiomics a suitable tool for screening for
PDAC. Through this, it may be possible to detect disease at an earlier stage when a larger proportion of
patients are amenable to surgical resection.

Pancreatic cysts present a diagnostic challenge and comprise a heterogeneous group of lesions with
biological behavior varying between benign indolent lesions and a propensity for progression to invasive
cancer. Radiomics has been applied to differentiate cyst types. In one of these studies, the authors were able
to differentiate serous cystadenomas and pancreatic cystic lesions with an AUC of 0.77 based on 22
[42]
radiomic features, which outperformed clinical and standard imaging features (AUC: 0.71) . In another
study, Xie et al. were able to differentiate macrocystic SCNs and MCNs with an AUC of 0.99 . Available
[43]
literature on studies applying radiomics to pancreatic cystic lesions was summarized by Machicado et al.,
who reported AUCs between 0.77 and 0.99 in differentiating different cyst types . Mucinous cystic lesions
[44]
of the pancreas are precursors to pancreatic cancer, but not all of these patients will go on to develop
cancer . Therefore, risk stratification is essential if we are to resect lesions prior to their progression to
[10]
cancer while avoiding surgery in patients with benign lesions given the high morbidity and mortality
associated with these procedures. Radiomics has been applied to the characterization of these cysts and has
shown an accuracy of 84% in distinguishing between common types of pancreatic cysts .
[45]

Currently, work is being performed on assessing the utility of radiomics in detecting high-grade dysplasia in
these lesions and predicting the risk of progression to pancreatic cancer. Tobaly et al. trained a radiomics
model based on preoperative CT to differentiate low-grade dysplasia, high-grade dysplasia, and invasive
cancer in patients with IPMN and demonstrated an AUC of 0.84 and 0.71 on internal and external
validation, respectively . Similarly, Polk et al. combined radiomics with conventional variables (thickened
[46]
and enhanced cyst wall and enhanced mural nodule) and reported an AUC of 0.93 (95%CI: 0.85-1.0) in
differentiating low-grade dysplasia and high-grade dysplasia or invasive cancer . If these tools are
[47]
developed, we will be able to accurately screen high-risk patients and recommend appropriate care.

Radiomic-based analysis could also improve surgical planning. Determination of resectability and the
likelihood of margin negative resections, particularly in the setting of neoadjuvant therapy, is challenging,
and the ability of conventional pancreas protocol CT (PPCT) to determine this remains low [48,49] . This is
driven by the fact that it is difficult to differentiate dead tissue and viable tumors on a PPCT. Since
radiomics provides a greater deal of information regarding the tissue in the region of interest, studies have
applied radiomics to detect the presence of vascular invasion and predict positive resection margins and
demonstrated improved prediction compared to conventional PPCT [50-52] . Recently, Schlanger et al.
systematically reviewed studies employing artificial intelligence and machine learning across two categories: