Lesion Detection and Segmentation¶

The ground truth is based on human annotations of GTVp and GTVn for the segmentation task. The contours were manually delineated by an expert radiation oncologist and underwent quality control by a board of 4 experts (certified as both radiologists and/or nuclear medicine physicians) to ensure consistency across all datasets. Precise contouring guidelines were elaborated to ensure the unification of all annotations. We used quality control methodology to ensure homogeneity and quality of contours across all centers.

Primary tumor delineation guidelines (GTVp)¶

Oropharyngeal lesions are contoured on PET/CT using information from PET and unenhanced CT acquisitions. The contouring includes the entire edges of the morphologic anomaly as depicted on unenhanced CT (mainly visualized as a mass effect) and the corresponding hypermetabolic volume, using PET acquisition, unenhanced CT, and PET/CT fusion visualizations based on automatic co-registration.

The contouring excludes the hypermetabolic activity projecting outside the physical limits of the lesion (for example, in the lumen of the airway or on the bony structures with no morphologic evidence of local invasion).

Standardized nomenclature according to AAPM TG-263: GTVp.

Special situations:

• Check clinical nodal category to make sure nearby FDG-avid and/or enlarged lymph nodes (e.g., submandibular, high level II, and retropharyngeal) are excluded.
• Clinical data was reviewed to rule out pre-radiation tonsillectomy or extensive biopsy in the case of the tonsillar fossa or base of tongue fullness/enlargement without corresponding FDG avidity. If so, the case was excluded.

ROI numbering:

• When more than one volume (rare): GTVt1, GTVt2, …
• When none (5% of cases [Kennel et al. 2019]), no region should be created.

Metastatic lymph nodes delineation guidelines (GTVn)¶

Lymph nodes are contoured on PET/CT using information from PET and unenhanced CT acquisitions. The contouring includes the entire edges of the morphologic lymphadenopathy as depicted on unenhanced CT and the corresponding hypermetabolic volume, using PET acquisition, unenhanced CT, and PET/CT fusion visualizations based on automatic co-registration for all cervical lymph node levels.

Standardized nomenclature for lymph node ROI: GTVn.

The contouring excludes the hypermetabolic activity projecting outside the physical limits of the lesion (for example, on the bordering bony, muscular, or vascular structures).

ROI numbering:

• When more than one: GTVn01, GTVn02, ... or all contours in a single label GTVn
• When none, no region should be created.

Limit on size and SUV: Pathologically confirmed OR SUV>2.5 OR diameter >=1cm, irrespective of the number of nodes.

Separation of GTVns: If several GTVns are "merged"/"touch": keep one structure with all of them. GTVn and GTVp must be separated.

TN Staging Classification¶

For the TN staging classification, the ground truth is the T and N stages for the patients. The T stages are T1, T2, T3, and T4, and the N stages are N0, N1, N2, and N3.

The original N stages coming from the hospital data includes misalignments between the N stages and the associated segmentation masks. Thus, the N stage information provided in the clinical data for the HECKTOR 2026 challenge underwent label refinment. The process of label refinment included regenerating the N labels according to the AJCC seventh edition staging guidelines.

Recurrence-Free Survival Prediction¶

For the outcome prediction task, the selected clinical endpoint is Recurrence-Free Survival (RFS), defined as time without any recurrence, censoring all others, including deaths. In particular, local, regional, and distant metastases are events and all others are censored. Time to event, defined in days, starts with the end of radiation therapy. All patients receive curative treatment and live with no cancer until recurrence.

REFERENCES¶

[Kennel et al. 2019] Kennel T, et al. "Head and neck carcinoma of unknown primary", European Annals of Otorhinolaryngology, Head and Neck Diseases, 136(3): 185-192 (2019).