HERMES: A Hybrid Ensemble for Head-and-Neck Tumor Segmentation, TN Staging, and Recurrence-Free Survival on PET/CT

2026-07-29

AI Summary

This study addresses the integrated challenge of automatic head and neck tumor segmentation, TN staging, and recurrence-free survival (RFS) prediction by proposing a unified containerized framework leveraging FDG-PET/CT imaging and electronic health records. The approach combines a 10-fold ensemble of STU-Net Small for segmentation, geometric feature–driven TN staging, and a tailored concordance-tracking survival loss function, enhanced by a decorrelated dual-submission strategy to improve staging accuracy under low-dimensional feature constraints. Through an equally weighted ensemble of deep learning and clinical risk experts, the method achieved a weighted score of 0.6454 on the HECKTOR 2026 validation leaderboard—comprising a Dice score of 0.641, T-stage accuracy of 0.580, N-stage accuracy of 0.642, and an RFS C-index of 0.679—securing advancement to the test phase.

Abstract

We present HERMES (Hybrid Ensemble for Radiotherapy-target segmentation, Malignancy staging, and Event-free Survival), a single containerized algorithm for the three HECKTOR 2026 subtasks: segmentation of the primary tumor (GTVp) and pathological lymph nodes (GTVn), radiological T/N staging, and recurrence-free survival (RFS), computed from a paired FDG-PET/CT scan and an electronic health record. A 10-fold ensemble of STU-Net Small networks produces the segmentation; the predicted mask then drives two downstream tasks. Rather than pass a generic radiomics vector to the staging models, we derive from the predicted masks a compact set of geometry features aligned with the size and number axes of AJCC/UICC 7th-edition radiological N/T staging. On internal cross-validation these features raise N-stage balanced accuracy from 0.691 to 0.720 (+0.030), our largest single design gain, at lower feature dimensionality. For prognosis we combine complementary deep and clinical risk experts in an equal-weight ensemble, and train one deep expert with a concordance-tracking survival loss of our own, whose value approximates the concordance index during training. Every component was selected on honest out-of-fold predictions under a regularization-oriented protocol, with no tuning on the public validation set, and deployed as two decorrelated submissions. On the HECKTOR 2026 validation leaderboard, HERMES achieved a weighted score of 0.6454 (Mean Dice 0.641, T balanced accuracy 0.580, N balanced accuracy 0.642, RFS C-index 0.679) and qualified for the testing phase. Team: AMC_HNC.

Problem

• tumor segmentation
• TN staging
• recurrence-free survival
• PET/CT
• head-and-neck cancer

Innovation

• hybrid ensemble
• geometry-aware radiomics
• concordance-tracking loss
• AJCC/UICC-aligned staging
• multi-task PET/CT analysis