MACHINE LEARNING ANALYSIS OF ENDOSCOPIC ULTRASONOGRAPHY TEXTURE AND CLINICAL INFORMATION PREDICTS PANCREATIC NEOPLASTIC PROGRESSION IN HIGH-RISK INDIVIDUALS WITHIN 18 MONTHS

Background:
Current abdominal imaging techniques used for pancreas ductal adenocarcinoma (PDAC) screening of high-risk individuals (HRI) such as endoscopic ultrasound (EUS), cannot accurately predict neoplastic progression to high-grade Pancreatic Intraepithelial Neoplasia/ Intraductal Papillary Mucinous Neoplasm (PanIN/IPMN) or PDAC. We aimed to develop a machine learning system that combines EUS features with high-value clinical features to predict neoplastic progression within 18 months.
Methods:
We used the data from the Cancer of Pancreas Screening Study, which is a prospective cohort of HRIs undergoing surveillance for PDAC at Johns Hopkins Hospital. HRIs consisted of asymptomatic unaffected individuals from familial PDAC kindreds or carriers of germline genetic pathogenic variants. We included a convenience sample of patients who underwent surgical resection of the pancreas within 18 months of the last EUS for suspected neoplasm. We categorized patients according to the highest grade of neoplasia in surgical pathologic specimens into those with low-grade neoplasia (low-grade PanIN/IPMN) and those with high-grade neoplasia (high-grade PanIN/IPMN or PDAC). We used high-value clinical features consisting of patient demographics, family history, medical history, and genetics. We quantified texture in each EUS image using 75 radiomic metrics to identify the features predicting the progression to HGD or PDAC. We computed gray-level co-occurrence matrices using pixel intensities within manually annotated regions of interest in pancreas parenchyma excluding macroscopic lesions. We extracted six features from the matrices that captured heterogeneity across the image. We trained three separate models (with clinical features alone, EUS texture metrics, and combined clinical and EUS features) using Adaptive Boosting, an ensemble machine learning method, to predict neoplastic progression. We evaluated the algorithms using split sample cross-validation. We estimated accuracy, AUC, sensitivity, and specificity.
Results:
We included 40 HRIs of whom 19 (47.5%) had low-grade neoplasia on surgical resection pathology while the other 21 (52.5%) had HGD or PDAC. The mean age at diagnosis was 66 years, 50% were female and 100% were Caucasians (Figure 1). Using 24 clinical features, the model had 50% accuracy in predicting progression to HGD or PDAC (Figure 2). EUS texture analysis-based model using 726 images had 78% accuracy. The model using combined clinical and EUS features had 96% accuracy.
Conclusions:
A machine learning model using EUS texture features and clinical features can accurately predict neoplastic progression within 18 months in HRIs. This will potentially aid in the early and accurate detection of pancreas cancer but requires further validation in a larger study before widespread implementation.