RADCURE Head and Neck Cancer RT Dataset

Overview

RADCURE is a large open-source head and neck cancer CT dataset comprising 3346 patients treated with curative-intent radiotherapy at Princess Margaret Cancer Centre (Toronto, Canada). The dataset includes RT simulation CT scans, manually generated and QA-reviewed target and organ-at-risk contours in DICOM RT-STRUCT format, and longitudinal clinical metadata. It is designed as a resource for radiomics, auto-segmentation, and treatment outcome prediction research. Hosted on The Cancer Imaging Archive (TCIA) and published by Welch et al. 2024. PMID:38362943

Composition

  • Cancer type: HNSC (head and neck squamous cell carcinoma), n=3346.
  • Disease subsites: oropharyngeal 50%, laryngeal 25%, nasopharyngeal 12%, hypopharyngeal 5%.
  • Demographics: median age 63; 80% male.
  • Modality: RT simulation CT acquired on systems from three manufacturers under standard clinical protocols.
  • Annotations: manually generated target volumes (gross primary tumour, gross nodal volumes) and 19 organs-at-risk contours, reviewed at weekly RT quality assurance rounds; standardised nomenclature applied.
  • Clinical metadata: demographic, clinical, and treatment information; staging per 7th edition TNM.
  • Follow-up: median 5 years; 60% surviving at last follow-up.
  • Format: images and contours as DICOM CT and RT-STRUCT; clinical data as CSV. PMID:38362943

Assays / panels (linked)

Papers using this cohort

  • PMID:38362943 — Welch et al. 2024, Medical Physics: primary dataset descriptor; 3346 HNC RT planning CTs from Princess Margaret Cancer Centre with standardised RT-STRUCT contours and longitudinal outcomes.
  • PMID:37397861 — Kim et al. 2023, Radiotherapy and Oncology: RADCURE used as the primary training (n=1,802) and internal test (n=750) cohort in the HNSC prognostic-modeling challenge; split by diagnosis date (2005–2013 vs 2016–2018).

Notable findings derived from this cohort

  • A custom data-mining and processing system was built to extract imaging and structure-set data from the institution’s RT planning and oncology information systems, linking each scan to longitudinal clinical outcomes. PMID:38362943
  • Standardised RT-STRUCT nomenclature was applied across 3346 patients to improve interoperability for downstream radiomics and auto-segmentation research. PMID:38362943
  • In the multi-institutional prognostic challenge, the top-performing model (MTLR + EMR + tumor volume) achieved AUROC = 0.823 on the RADCURE internal test set; adding deep radiomics to EMR did not improve AUROC. PMID:37397861

Sources

  • TCIA collection: RADCURE — https://www.cancerimagingarchive.net/collection/radcure/
  • PMID:38362943 — Welch et al. 2024, Medical Physics, DOI 10.1002/mp.16971.
  • PMID:37397861 — Kim et al. 2023, Radiotherapy and Oncology.

This page was processed by crosslinker on 2026-05-04.