Head and Neck Tumor Segmentation for MRI-Guided Applications Challenge

Radiation therapy (RT) is a cornerstone of cancer treatment for a wide variety of malignancies. Chief among the beneficiaries of RT as a treatment modality is head and neck cancer (HNC). Recent years have seen an increasing interest in MRI-guided RT planning. As opposed to more traditional CT-based RT planning, MRI-guided approaches afford superior soft tissue contrast and resolution, allow for functional imaging through special multiparametric sequences (e.g., diffusion-weighted imaging [DWI]), and permit daily adaptive RT through intra-therapy imaging using MRI-Linac devices. Subsequently, improved treatment planning through MRI-guided adaptive RT approaches would maximize tumor destruction while minimizing side effects. Given the great potential for MRI-guided adaptive RT planning, it is anticipated that these technologies will transform clinical practice paradigms for HNC. 
 
The extensive data volume for MRI-guided HNC RT planning makes manual tumor segmentation by physicians — the current clinical standard — often impractical due to time constraints. This is compounded by the fact that HNC tumors are among the most challenging structures for clinicians to segment. Artificial intelligence (AI) approaches that leverage RT data to improve patient treatment have been an exceptional area of interest for the research community in recent years. The use of deep learning in particular has made significant strides in HNC tumor auto-segmentation. These innovations have largely been driven by MICCAI public data challenges such as the HECKTOR Challenge, and the SegRap Challenge. However, to-date, there exist no large publicly available AI-ready adaptive RT HNC datasets for public distribution. It stands to reason that community-driven AI innovations would be a remarkable asset to developing technologies for the clinical translation of MRI-guided RT. 
 
In this proposal, we describe a data science challenge focused on the segmentation of HNC tumors for MRI-guided adaptive RT applications. The challenge will be composed of 2 tasks focused on automated segmentation of tumor volumes on 1. pre-RT multiparametric MRI images and 2. mid-RT multiparametric MRI images. Our challenge is particularly unique as it seeks to ascertain whether incorporating prior timepoint data into auto-segmentation algorithms leads to enhanced performance for RT applications.