Uncertainty-guided pancreatic tumor auto-segmentation with Tversky ensemble

Area of Science:

• Medical imaging and artificial intelligence
• Radiomics and computational pathology
• Oncology and surgical planning

Background:

• Pancreatic gross tumor volume (GTV) segmentation is complex due to variable morphology and uncertain ground truth.
• Existing deep learning methods face challenges with uncertain ground truth and lack customization options.
• Accurate GTV delineation is crucial for effective pancreatic cancer treatment planning.

Purpose of the Study:

• To develop a human-in-the-loop tool for pancreatic GTV segmentation.
• To leverage Tversky ensembles and uncertainty estimation for improved segmentation accuracy.
• To provide clinicians with customizable segmentation outputs for enhanced confidence.

Main Methods:

• Utilized 282 patient cases from the Medical Segmentation Decathlon (pancreas task).
• Trained a five-member segmentation ensemble using a Tversky loss layer.
• Employed uncertainty estimation and probability thresholding to generate final contours.
• Evaluated eleven contours against ground truths using Dice Similarity Coefficient (DSC), Hausdorff Distance 95 (HD95), and Mean Surface Distance (MSD).

Main Results:

• The Tversky ensemble achieved a DSC of 0.47, HD95 of 12.70 mm, and MSD of 3.24 mm.
• The proposed method outperformed the Swin-UNETR configuration, which previously held the state-of-the-art result.
• Optimal thresholding configurations were identified for generating accurate segmentation contours.

Conclusions:

• Ensemble-based uncertainty estimation is effective for pancreatic tumor segmentation.
• The tool offers a consensus probability map for clinician fine-tuning and customization.
• This approach enhances confidence in generating pancreatic GTV contours.