Pancreatic Tumor Segmentation in Therapeutic and Diagnostic MRI

Pancreatic cancer is among the most aggressive and lethal malignancies, with a 5 year survival rate of only 10-12%. The poor prognosis is largely due to late detection, rapid disease progression, and the difficulty of complete surgical resection. Medical imaging plays a central role in improving patient outcomes by enabling early detection, accurate staging, and precise treatment planning.

Magnetic Resonance Imaging (MRI) is particularly valuable for pancreatic cancer management due to its superior soft tissue contrast. This capability allows for detailed visualization of tumors, adjacent organs, and neurovascular structures—essential for assessing resectability and guiding treatment strategies, such as surgical resection or
neoadjuvant therapies. MRI is also critical in radiation therapy planning, where accurate delineation of tumors and organs at risk improves treatment precision and patient outcomes, especially in MR-Linac systems. These systems integrate real time MRI with radiation delivery, enabling adaptive treatments. However, precise and efficient segmentation of the pancreas and tumors remains a bottleneck in clinical workflows.

Manual segmentation of pancreatic tumors is time consuming and labor intensive, particularly when transitioning between diagnostic MRIs (high quality imaging used for initial assessment) and MR-Linac images (real time treatment imaging). Annotating MR-Linac data is further complicated by lower image quality and modality differences. The current workflow often involves transferring delineations from diagnostic scans to treatment scans using registration algorithms, a process that introduces delays and potential inaccuracies that need to be manually corrected. This challenge stems from the scarcity of annotated MR-Linac datasets, which difficult the development of robust AI segmentation models tailored to this modality.

The purpose of this challenge is to advance cross-domain learning for pancreatic cancer segmentation. By leveraging annotated diagnostic MRIs alongside a limited number of annotated MR-Linac images, participants will develop models capable of accurate and robust segmentation across both domains. A key goal is to reduce reliance on manual delineations and registration workflows, ultimately accelerating treatment planning and enhancing clinical efficiency.

The challenge will provide a training set comprising annotated diagnostic MRIs, unlabeled multi-sequence MRIs, and a smaller set of annotated MR-Linac images. Separate test sets for diagnostic MRIs and MR-Linac images, ensuring robust evaluation of performance.