Automated Prediction of Postoperative Pancreatic Fistula Using Preoperative Computed Tomography

Technical Summary: Automated Prediction of Postoperative Pancreatic Fistula Using Preoperative Computed Tomography

Problem Statement
Postoperative pancreatic fistula (POPF) is a severe complication following pancreatic resection, significantly increasing morbidity, hospital length of stay, and healthcare costs. The incidence of clinically relevant POPF (CR-POPF) ranges from 10% to 40%. While traditional risk stratification relies on clinical factors (e.g., BMI, pancreatic duct diameter) and intraoperative assessments, these methods often suffer from moderate accuracy and subjectivity. The authors posit that preoperative computed tomography (CT) scans contain rich, high-dimensional data—including subtle textural features and anatomical variations—that could be leveraged to better characterize risk before surgery, yet this potential remains underutilized by current statistical models.

Methodology
The paper proposes an end-to-end deep learning pipeline designed to predict the probability of CR-POPF using only preoperative CT imaging. The workflow consists of four primary stages:

1. Data Cohort and Preprocessing:

   • The study utilized a prospectively maintained dataset from Mayo Clinic comprising adult patients who underwent pancreatoduodenectomy between 2010 and 2020.
   • Two distinct cohorts were analyzed based on contrast phase: Portal Venous (N=1,353) and Late Arterial (N=963).
   • Phase Detection: An XGBoost classifier determined the contrast phase of each scan based on Hounsfield Unit (HU) intensities.
   • Segmentation and ROI Extraction: To focus the model on relevant anatomy, the authors employed automated pancreas segmentation to generate binary masks. Two segmentation sources were compared: a custom Mayo Clinic 3D nnU-Net model (MS) and the TotalSegmentator model (TS).
   • Standardization: The pipeline applied HU windowing (window width=300, center=80) to standardize intensities. Volumes were cropped to a tight bounding box around the pancreas and center-padded to a uniform size ($H^\star_{in} \times W^\star_{in} \times D^\star_{in}$).

2. Model Architectures:
   Three 3D convolutional neural network (CNN) architectures were evaluated for binary classification (POPF present vs. absent):

   • CNN3D: A custom, lightweight baseline network trained from scratch, featuring three convolutional stages with residual shortcuts.
   • ResNet-(2+1)D-18: A residual architecture that factorizes 3D convolutions into spatial and depth (inter-slice) convolutions to efficiently model 3D context.
   • ResNet-MC3-18: A model utilizing mixed spatiotemporal convolutions, employing 3D kernels in early layers for inter-slice context and slice-wise spatial convolutions in later layers.

3. Training and Evaluation:

   • Augmentation: Training utilized stochastic MONAI transforms, including random 3D flips, rotations, zooms, and intensity perturbations (histogram adjustments, noise, smoothing).
   • Optimization: Models were trained using AdamW with a OneCycle learning rate schedule. Class imbalance was addressed via weighted sampling and weighted cross-entropy loss.
   • Validation: Performance was assessed using 5-fold stratified cross-validation under two protocols: (i) balanced train/test splits and (ii) natural distribution splits preserving the observed CR-POPF prevalence.

Key Results
The study evaluated performance using Area Under the Receiver Operating Characteristic Curve (AUC), accuracy, precision, recall, and F1 scores.

• Architecture Performance: Residual spatiotemporal backbones (ResNet-(2+1)D-18 and ResNet-MC3-18) consistently outperformed the custom CNN3D baseline.
• Segmentation Impact: The quality of the pancreas mask significantly influenced results. Models using the Mayo Segmentation (MS) masks achieved higher AUCs (approx. 0.73) compared to those using TotalSegmentator (TS) masks (approx. 0.65–0.66), highlighting the importance of ROI fidelity.
• Contrast Phase: In the Late Arterial cohort, the best-performing model (R2Plus1DResNet18 with MS masks) achieved an AUC of $0.73 \pm 0.04$ in balanced splits and $0.68 \pm 0.07$ in natural distribution splits. In the Portal Venous cohort, R2Plus1DResNet18 with MS masks achieved an AUC of $0.73 \pm 0.06$ (balanced) and $0.70 \pm 0.03$ (natural).
• Class Imbalance: Under natural distribution splits, precision decreased while recall remained high, a typical behavior for imbalanced medical datasets, though the model ranking by AUC remained consistent.

Significance and Claims
The authors claim that this work demonstrates the feasibility of using preoperative CT imaging alone to stratify patients for CR-POPF risk. Key contributions include:

• Methodological Benchmark: The paper provides a rigorous comparison of modern 3D deep learning architectures for pancreas-specific CT classification, establishing a baseline for future research.
• Clinical Utility: The proposed pipeline offers a potential tool for preoperative decision-making. By identifying high-risk patients prior to surgery, clinicians could theoretically tailor perioperative management, consider alternative surgical approaches (e.g., avoiding anastomosis), or institute enhanced postoperative monitoring.
• Data-Driven Insights: The results suggest that pancreas-centric, HU-windowed preoperative CT scans contain measurable imaging signatures associated with CR-POPF that are amenable to computational extraction.

Limitations
The authors explicitly note several limitations:

• Scope of Prediction: Imaging alone cannot fully determine fistula development, as the outcome is also influenced by patient-level factors (age, comorbidities) and intraoperative variables (blood loss) not included in this model.
• Error Propagation: The pipeline relies on automated segmentation; imperfect masks or phase misclassification could propagate errors to the final prediction.
• Validation: The study lacks external validation on independent multi-institution cohorts and prospective evaluation, which are necessary to confirm robustness and clinical utility before deployment.