Page 61 - Read Online

Page 19 Tovar et al. Art Int Surg 2023;3:14-26 https://dx.doi.org/10.20517/ais.2022.38

coupling of organ-specific radiomics, which requires organs to be identified within the images for analysis,
and AI [Table 2], which can provide both organ labels and integrated analysis of the radiomics features, may
be a powerful tool for analyzing routine clinical images alongside a radiologist to provide new clinical
insights such as a prediction of malignancy [Figure 1] [56,57] .

Several methods utilized in recent research have incorporated image features seen in multiple image
modalities, such as computed tomography (CT) and magnetic resonance imaging (MRI), to create models
to detect those with a higher chance of malignancy. Qureshi et al. conducted a retrospective study of 72
subjects to analyze images to find precursor indicators of PDAC present in pre-diagnostic CT scans, which
were taken 3-6 months prior to PDAC diagnosis when indicated as normal by a pathologist. Their Bayes
classifier model detected image features able to categorize scans as ‘healthy’ or ‘pre-diagnostic’ with an
[41]
accuracy of 86% .
Chen et al. investigated the use of a multi-state model of abnormal pancreatic morphological features from
CT and MRI in combination with patient demographics, clinical features, and lab measurements for risk
prediction. Out of the PDAC abnormalities evaluated, the most prevalent were pancreatic parenchymal
atrophy reported in 21.4% of patients and calcification in 12.6% of patients. Among these morphological
features, pancreatic duct dilatation was determined as an additional indicator of PDAC. The model found
that those with a calculated risk of more than 5% represented 90% of their total PDAC study population
[42]
(AUC 0.825-0.833) .

EARLY DETECTION MODELS
Radiomics-based AI models
[55]
Signs of pancreatic cancer have previously been estimated to be detectable 3-36 months before diagnosis .
Mukherjee et al. trained a ML model to detect PDAC at a stage not visible on CT imaging by radiologists.
Using a pre-diagnostic cohort of 155 patients and a control cohort of 265 patients, CT scans were manually
segmented, then radiomic CT features were extracted and selected. Four ML classifiers were trained, with
Support Vector Machine performing the best in classifying CT scans as ‘pre-diagnostic’ or ‘normal’ when
evaluating specificity, sensitivity, AUC, and accuracy (AUC, 0.98). All four ML models performed better
than the radiologists, who performed with an AUC of 0.66 . This indicates the promising potential to use
[43]
AI in conjunction with normal imaging to aid radiologists in detecting potential malignancy. Comparably,
another study sought to increase pancreatic detection in tumors smaller than 2 cm, which are often missed
[58]
by radiologists . Using a CNN-trained model, the pancreas and tumor were segmented from contrast-
enhanced CT scans. This DL-based computer detection model was used in 546 patients with pancreatic
cancer and a control group of 733 in Taiwan .
[50]
One area of interest is the ability of radiomics to predict malignancy risk in patients with cystic IPMNs,
which can transform to PDAC. IPMNs arise from the pancreatic duct and side branches and are estimated
to account for approximately 10% of PDAC patients. Notably, 3% of the general population is estimated to
have an IPMN, indicating that many of these lesions are benign . Predicting whether these neoplasms are
[59]
malignant on imaging can be a valuable tool in early detection; however, current imaging assessment is
challenging and not accurate in predicting malignancy risk. Still, the current Fukuoka International
Consensus Guidelines (ICG) (and other guidelines) use morphological imaging features to guide the
decision to proceed with surgical resection . This carries a risk of overtreatment, since pancreatectomy is
[60]
associated with the highest rates of morbidity (40%) and mortality (up to 2%) among abdominal
surgeries . Hanania et al. previously showed the correlation between radiomic features and
[61]
histopathological grade of IPMNs. In their logistic regression model, an AUC of 0.96 was achieved in