hecktor 2025 report

baixiang zhao, Surajit Ray

Keywords: radiotherapy planning, multi-modality neural network, survival analysis.

TL;DR: This is the technical report of the AI_state team for the HECKTOR 2025 challenge, covering all three tasks.

Abstract:

In this paper, we present our submission to the HECKTOR
2025 challenge, addressing all three tasks using PET and CT imaging
together with electronic health record (EHR) data. The tasks include
segmentation of primary tumors and lymph nodes, survival prediction,
and HPV-status classification.For the first task (segmentation), we used
a U-Net style SegResNet and achieved a Dice score of 0.52 for the primary
tumor volume and an aggregated Dice score of 0.38 for lymph node volume
on the validation set. Our performance in this task secured us a Top
10 ranking among all participating teams.For the second task (survival
prediction), we designed a multimodal model that combines imaging features
with clinical data, reaching a C-index of 0.6482 on the validation
cohort, also placing us in the Top 10 teams.We applied the same multimodal
model to the third task (HPV-status classification), which resulted
in a balanced accuracy of 0.4655, achieving 2nd place in this specific subchallenge.
Our experiments show that combining features across multiple
data folds improves performance, while imbalanced classes reduce accuracy.
We also note that adding more information, such as radiotherapy
planning data and tumor volume, which may be important for prognosis,
could further improve results in future work.

Submission Number: 5