Mirmojarabian, Mohammadali (2024) Preoperative CT-derived body composition as predictor of postoperative pancreatic fistula risk after whipple surgery. Masters thesis, Memorial University of Newfoundland.
Full text not available from this repository.
Abstract
Pancreaticoduodenectomy (PD) is a primary treatment for pancreatic cancer but carries a high risk of postoperative pancreatic fistula (POPF). Accurate prediction of POPF is key for improving patient outcomes and guiding surgical decisions. This study analyzed body composition (tissues within the lumbar range) and clinical factors for predicting POPF, using a cohort of 777 patients. Volumes (V) and Hounsfield Unit (HU) values were extracted from Computed Tomography (CT) scans using the Data Analysis Facilitation Suite (DAFS) CT segmentation software, and logistic regression was applied to analyze high- and low-risk groups (Fistula ±). For males, Age at Surgery (p = 0.021), while for females, the Charlson Comorbidity Index (CCI) (p = 0.018) and Malignancy (p = 0.037) were significantly associated with Fistula ±. In males, Liver HU (p = 0.037), Gallbladder HU (p = 0.037), and Heart HU (p = 0.040), while in females, Lung V (p = 0.044), Lung HU (p = 0.002), Trachea V (p = 0.029), Trachea HU (p = 0.017), Heart V (p = 0.027), and Aorta V (p = 0.001), and for both males and females respectively, Pancreas V (p = 0.006 and 0.047) exhibited significant differences with Fistula ±. For tissue features, in males, SKM V (p = 0.029), while in females, SKM HU (p = 0.006), FAT HU (p = 0.006), VAT HU (p = 0.000), SAT V (p = 0.032), and SAT HU (p = 0.012), and in both genders, FAT V (p = 0.012 and 0.021), VAT V (p = 0.002 and 0.044), and the VAT/SKM V ratio (p = 0.006 and 0.032) were significantly associated with Fistula ±. For prediction, a sub-cohort of 140 scans was used, with 11 key features selected. A weighted stratified five-fold cross-validation approach was applied, with SHapley Additive exPlanations (SHAP) ranking Age at Surgery, Pancreas HU, Liver HU, Aorta V, FAT HU, and SKM HU as top predictors. The model achieved a mean Area Under the Curve (AUC) of 0.85 (training) and 0.80 (testing), with a mean Balanced Accuracy of 0.80 and 0.78, respectively. These results highlight preoperative body composition as a non-invasive, effective tool for predicting POPF risk and optimizing surgical planning, while underscoring the importance of personalized, gender-specific risk assessments.
| Item Type: | Thesis (Masters) |
|---|---|
| URI: | http://research.library.mun.ca/id/eprint/16847 |
| Item ID: | 16847 |
| Additional Information: | Includes bibliographical references (pages 75-97) -- Restricted until December 20, 2025 |
| Keywords: | postoperative pancreatic fistula (POPF), medical imaging and artificial intelligence, statistical significance analysis, machine learning risk prediction modeling, full body composition analysis |
| Department(s): | Science, Faculty of > Computer Science |
| Date: | September 2024 |
| Date Type: | Submission |
| Library of Congress Subject Heading: | Pancreatic fistula; Pancreaticoduodenectomy; Cancer--Surgery--Complications; Body composition--Analysis; Cancer--Imaging; Machine learning; Artificial intelligence |
Actions (login required)
 |
View Item |
