Get started
Work

GASP: Trained model weights for Multimodal Geometry-Aware Segmentation, Staging, and Prognostication of Head and Neck Cancer (MICCAI 2026 HECKTOR Challenge)

Year: 2026
Type: software
Abstract: Trained model weights for GASP, a geometry-aware framework for the MICCAI 2026 HECKTOR challenge, performing multimodal segmentation, staging, and prognostication of head and neck cancer from co-registered FDG-PET and CT. The bundle contains a five-fold nnU-Net 3D full-resolution ensemble for segmentation of the primary tumor (GTVp) and nodal disease (GTVn), together with the lightweight downstream models: a size-focused logistic classifier for nodal (N) stage, a label-audited logistic classifier for tumor (T) stage, and a Cox proportional hazards model for recurrence-free survival. On the official challenge test set of 400 patients withheld from participants, the system achieved a mean Dice of 0.652 for segmentation, a recurrence-free survival concordance index of 0.558, and balanced accuracies of 0.461 and 0.564 for tumor and nodal staging, giving an overall weighted score of 0.5654 under the HECKTOR aggregation of 0.25 for segmentation, 0.35 for TN staging, and 0.40 for survival. Contents:- nnUNet_results/: five fold checkpoints for Dataset501_HECKTOR (nnUNetTrainer_250epochs, 3d_fullres)- n_classifier.joblib: nodal staging model- t_classifier.joblib: tumor staging model- staging_Tmodel.joblib: earlier tumor staging model retained for compatibility- prognosis_model.joblib: Cox survival model- staging_config.json: configuration used by the inference container Code and usage instructions: https://github.com/sanketkachole/MICCAI26_Challenge_HECKTOR The HECKTOR imaging data is not included and must be obtained from the challenge organisers under their data use agreement.
DOI: 10.5281/zenodo.22073791
Open Access status: green
OpenAlex ID: W7204124407
Cites:
Cited by:
Related to: