PANTHER Challenge Report: Cross-Domain Pancreatic Tumor Segmentation in Magnetic Resonance Imaging

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Accurate pancreatic tumor delineation on MRI is crucial for diagnosis, radiotherapy planning, and outcome assessment.
• Manual contouring is time-consuming and prone to inter-observer variability.
• Existing benchmarks for pancreas tumor segmentation primarily use CT, not MRI.

Purpose of the Study:

• To establish the first public benchmark for automatic pancreatic tumor segmentation on MRI.
• To evaluate automated segmentation performance on both diagnostic MRI and MRI-Linac systems.
• To address the gap in MRI-based pancreas tumor segmentation benchmarks.

Main Methods:

• The Pancreatic Tumor Segmentation in Therapeutic and Diagnostic MRI (PANTHER) challenge was organized.
• A dataset with expert annotations for diagnostic MRI and MRI-Linac scans was created.
• Performance was assessed using overlap, distance, and volume error metrics.

Main Results:

• Top methods achieved performance comparable to inter-reader agreement on diagnostic MRI.
• Automated models on diagnostic MRI showed potential for clinical utility.
• Segmentation performance on MRI-Linac images was significantly lower and more variable, with localization failures.

Conclusions:

• Clinically useful automated pancreatic tumor segmentation is achievable on diagnostic MRI.
• Robust Gross Tumor Volume (GTV) segmentation on MRI-Linac remains an open challenge.
• Annotation quality and consensus are critical for training effective segmentation models.