SoMNet：Automatic Contouring of Head and Neck Organs at Risk Guided by Boundary Information

To enhance the application of radiotherapy planning, we developed a novel head and neck imaging dataset, HND, comprising simulation X-ray computed tomography (CT) images from 486 patients with head and neck cancers who underwent intensity-modulated radiotherapy (IMRT) between February 2019 and June 2024 (Table 1). The inclusion criteria were as follows: (1) the CT scan range must encompass both the head and neck, with the upper boundary extending at least 3 cm above the skull base line and the lower boundary reaching 2–3 cm below the inferior margin of the clavicular heads; (2) the simulation CT images must be clear and free of metal artifacts. All CT images were acquired using a SIEMENS CT scanner with a tube voltage of 140 kV, a current of 280 mAs, a slice thickness of 3 mm, and a pitch of 1:1. The pixel spacing was standardized at 0.98 mm, without appearance enhancement. Each CT volume consisted of 83 to 210 slices with a resolution of 512 × 512 pixels per slice. The in-plane resolution ranged from 0.795 mm × 0.795 mm to 1.563 mm × 1.563 mm, with an interslice spacing of 3.0 mm, reflecting the high resolution of the HND dataset. This study was approved by the Institutional Review Board of Shandong Cancer Hospital. All images were anonymized to ensure patient privacy, and all clinical treatment details were removed.

The HND dataset was randomly divided into three subsets for training, validation, and testing, comprising 341, 97, and 48 cases, respectively. Thirty OARs in the CT images were manually delineated slice by slice using 3D-Slicer (v5.6.2) by a senior oncologist with 15 years of experience. Subsequently, these annotations were meticulously reviewed and revised by an oncology expert with over 20 years of experience. In cases of discrepancies, discussions were conducted to reach a consensus, ensuring the accuracy and reliability of the annotations. These final consensus annotations were released to support the development and evaluation of methodological or clinical applications. Importantly, all delineations and consensus discussions adhered to the contouring guidelines issued by the Radiation Therapy Oncology Group (RTOG) (https://www.rtog.org/). We analyzed the proportion of annotation inconsistencies between the senior oncologist and the expert. The delineated OARs included the brainstem, spinal cord, temporal lobe (left and right, L/R), optic nerve (L/R), optic chiasm, pituitary, mandible, temporomandibular joint (L/R), eye (L/R), lens (L/R), middle ear (L/R), inner ear (L/R), parotid gland (L/R), submandibular gland (L/R), oral cavity, larynx, esophagus, thyroid, pharynx, and cochlea (L/R).