Clinical Problem 🏥¶

Pancreatic cancer remains one of the most lethal malignancies, with a five‐year survival rate of only 10–12% due to late detection and rapid progression. Magnetic Resonance Imaging (MRI) plays a central role in its management, offering visualization that is critical for staging, treatment planning, and guiding adaptive radiation therapies, particularly with MR‐Linac systems. However, the visualization of pancreatic tumors is exceptionally challenging, requiring significant expertise from radiologists due to the pancreas’ complex anatomy and the often subtle appearance of tumors on MRI.

Furthermore, the lack of publicly available datasets for pancreatic cancer MRI severely limits the development and validation of Artificial Intelligence (AI)-driven solutions. Manual segmentation of both diagnostic and real time treatment images is time consuming and labor intensive, adding to the already high workload of clinical experts. The combination of these challenges highlights the urgent need for robust AI models capable of automating and improving the segmentation process, ultimately supporting clinicians in more efficient and accurate treatment planning.

Need for datasets and standardized benchmarks 📈¶

Despite the growing role AI in medical imaging, progress in pancreatic tumor segmentation on MRI remains severely constrained by the absence of publicly available datasets. This limitation is even more pronounced for MR-Linac images, where data scarcity hinders the development, validation, and benchmarking of AI-driven solutions. Without standardized datasets, algorithm performance cannot be fairly compared, limiting scientific progress and clinical adoption.

PANTHER is the first-ever challenge to address pancreatic tumor segmentation on MRI across different sequences and applications—including T1-weighted arterial contrast-enhanced MRI for diagnostics and T2-weighted MRI for MR-Linac adaptive radiotherapy. By introducing the first dataset for this task, PANTHER aims to bridge the gap between research and clinical practice, fostering the development of AI solutions that improve segmentation accuracy and streamline workflows in both diagnostic and treatment settings.

The PANTHER Challenge 👨🏼‍⚕️👩🏻‍💻¶

PANTHER is the first grand challenge dedicated to pancreatic tumor segmentation on MRI, addressing both diagnostic and treatment planning needs. The challenge provides a unique opportunity to develop and benchmark deep learning models for pancreatic cancer segmentation across different MRI sequences, helping bridge the gap between research and clinical implementation.

The challenge consists of two key tasks:

1. Pancreatic Tumor Segmentation in Diagnostic MRIs

   • This task focuses on automating tumor delineation on T1-weighted arterial contrast-enhanced MRI, a critical step for staging and treatment planning.
   • Participants will have access to 92 annotated images for supervised learning.
   • Additionally, 367 unannotated MRIs from different sequences (e.g., T1-weighted contrast-enhanced venous phase, diffusion-weighted imaging (DWI), T1-weighted in phase) will be provided, allowing teams to explore techniques such as unsupervised pretraining or self-supervised learning to enhance model performance.

2. Pancreatic Tumor Segmentation MR-Linac MRIs

   • This task focuses on segmenting pancreatic tumors on T2-weighted MRI for MR-Linac adaptive radiotherapy, a crucial step in real time treatment planning.
   • The dataset includes 50 annotated images, making this a real world few-shot learning problem.
   • Participants will need to explore transfer learning, data augmentation, or few-shot learning techniques to develop robust models despite the small dataset size.

By introducing the first publicly available dataset for this tasks, PANTHER aims to foster the development of AI solutions that can support clinicians in reducing manual segmentation workload, improving accuracy, and accelerating clinical workflows. This challenge not only presents a high impact, real world problem but also encourages participants to push the boundaries of modern deep learning by exploring state-of-the-art techniques such as unsupervised learning, few-shot learning, transfer learning, and self-supervised pretraining. By leveraging these approaches, participants can develop more robust and generalizable AI models that address the unique challenges of pancreatic tumor segmentation across different MRI sequences and clinical applications.

Prizes🏆¶

TASK 1

🥇1st place: € to be anounced

🥈2nd place: € to be anounced

🥉3rd place: € to be anounced

TASK 2

🥇1st place: € to be anounced

🥈2nd place: € to be anounced

🥉3rd place: € to be anounced

Publication Policy 📰¶

The PANTHER organizers will consolidate the challenge results and submit a comprehensive paper to a high-impact journal.

Up to three members from each of the top three performing teams per task will be invited as co-authors. If a team wishes to include more than three co-authors, they must notify the organizers and provide a justification for their inclusion. Additionally, teams whose submissions demonstrate notable scientific relevance or innovation may be invited to contribute to the publication.
Participants can independently publish methods based on the challenge data after an embargo period of 4 months from the challenge's final event.