2182. Diagnostic Potential of Machine Learning Analysis and Radiomics for Pancreatic Cysts
Authors * Denotes Presenting Author
  1. Adam Awe *; University of Wisconsin Hospital
  2. Michael Vanden Heuvel; University of Wisconsin
  3. Tianyuan Yuan; University of Wisconsin
  4. Mingren Shen; University of Wisconsin
  5. Victoria Rendell; University of Wisconsin Hospital
  6. Emily Winslow; Medstar Georgetown Transplant Institute
  7. Meghan Lubner; University of Wisconsin Hospital
Preoperative diagnosis of pancreatic cysts (PCs) is challenging. Texture analysis, an advanced radiomics tool that quantifies image heterogeneity, has demonstrated promise in enhancing diagnostic accuracy in non-pancreatic lesions (1-4). However, few studies have investigated the utility of texture and radiomics analysis in PCs (5-6). We aimed to identify radiomics features associated with mucinous PCs using a machine learning analysis.

Materials and Methods:
Surgical records at our institution from 1995-2017 were searched for resected PCs. Patients with non-pseudocyst PCs, surgical pathology, and contrast-enhanced computed tomography (CT) within 1 year prior to resection were included. Pancreatic cysts were segmented using a novel quantitative imaging software, HealthMyne. Texture analysis features were extracted from 3-dimensional lesion segmentations. Additional clinical information from electronic health records and radiologic features from the images were manually collected. A machine learning algorithm, extreme gradient boosting (XGBoost), that uses decision tree methods to prevent over-fitting was applied to the dataset to generate a classifier that delineates mucinous from non-mucinous PCs. A positive class weight scaling of 0.25 was used without oversampling parameters. Precision, recall, accuracy, and area under the curve (AUC) were generated by the XGBoost classifier. Additionally, composite metrics of precision and recall (F1-score), and sensitivity and specificity (Geometric mean; G-mean) were determined by the XGBoost classifier.

A total of 104 resected PCs were included. Surgical pathology identified 81 (78%) mucinous and 23 (22%) non-mucinous PCs. Mucinous PCs consisted of 25 (31%) mucinous cystic neoplasms, 53 (65%) intraductal papillary mucinous neoplasms, and 3 (4%) other. Non-mucinous PCs consisted of 17 (74%) serous cystadenomas, 2 (9%) lymphoepithelial cysts, and 4 (17%) other. The XGBoost mucinous classifier demonstrated a precision of 0.88 (± standard deviation; ±0.07), recall of 0.76 (±0.11), accuracy of 0.73 (±0.10), and AUC of 0.70 (±0.12). An F1-score and G-mean value were calculated at 0.81 (±0.08) and 0.67 (±0.15), respectively. The mucinous classifier determined metrics most predictive of mucinous PCs that included the square-root of the mean of Hounsfield units (HU), cyst location, mean HU, average row standard deviation of the grey-level co-occurrence matrix, and kurtosis.

Using the XGBoost algorithm to build a mucinous classifier of PCs demonstrated a good predictive accuracy evidenced by the F1-score with moderate sensitivity and specificity evidenced by the G-mean value. Texture features and general PC radiomics features on CT imaging were important in building a mucinous classifier. Applying machine learning principles to PC may provide non-invasive means of differentiating mucinous from non-mucinous cysts and deserve further consideration.