A Comparison of CT-Based Pancreatic Segmentation Deep Learning Models.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Radiology and Computational Pathology
• Biomedical Engineering and Machine Learning

Background:

• Accurate pancreas segmentation on CT scans is crucial for diagnosing pancreatic diseases and developing imaging biomarkers.
• Existing studies often rely on aggregate performance metrics, potentially masking variations in model performance across different patient and scan characteristics.

Purpose of the Study:

• To benchmark the performance of five leading pancreas segmentation models using multiple metrics.
• To evaluate how segmentation performance is affected by scan characteristics, including contrast status and peri-pancreatic attenuation.

Main Methods:

• A retrospective study identified five high-performing pancreas segmentation models (TotalSegmentator, Abdomen Atlas, nnUNetv1, AASwin, DM-UNet).
• Models were evaluated on 352 CT scans using Dice score, Hausdorff distance, and average surface distance.
• Results were stratified by contrast status and peri-pancreatic attenuation; multivariate regression identified factors associated with segmentation accuracy.

Main Results:

• TotalSegmentator, Abdomen Atlas, and AASwin were top performers with Dice scores around 77-80%.
• Performance decreased on non-contrast scans for AASwin and nnUNetv1 (P < .001).
• Increasing peri-pancreatic attenuation negatively impacted Dice scores for all models except TotalSegmentator (P < .01).

Conclusions:

• Convolutional neural network-based models trained on diverse datasets (TotalSegmentator, Abdomen Atlas, AASwin) showed the best performance.
• TotalSegmentator achieved comparable results to models trained on larger datasets, highlighting training data efficiency.
• Differential performance across patient and scanning characteristics necessitates comprehensive evaluation beyond aggregate metrics for clinical applicability.