Salivary glands auto-segmentation in head and neck radiotherapy: imaging modalities, models, and clinical relevance

Radiotherapy-induced salivary gland injury is a major contributor to xerostomia and impaired quality of life in patients with head and neck cancer, making accurate organ-at-risk delineation clinically important. Deep learning-based auto-segmentation has the potential to improve contouring efficiency and consistency, but reported performance is heterogeneous and its clinical significance is not always clear. This narrative review summarizes current evidence on salivary gland auto-segmentation across computed tomography (CT), magnetic resonance imaging (MRI), and prostate-specific membrane antigen positron emission tomography-computed tomography (PSMA PET-CT), with particular emphasis on gland-specific challenges, performance evaluation, and clinical translation. CT has the largest evidence base and remains most relevant to routine radiotherapy planning, whereas MRI provides improved soft-tissue visualization and PSMA PET-CT offers high functional contrast for selected salivary structures. Across modalities, segmentation performance is generally more challenging for smaller structures, particularly the sublingual and tubarial gland regions. Importantly, geometric metrics such as the Dice similarity coefficient do not alone establish clinical acceptability or dosimetric equivalence; evaluation should also consider boundary accuracy, expert correction burden, clinical acceptability, and dosimetric impact. Current evidence is further limited by heterogeneous study designs and reference standards, scarce external validation, and limited evidence linking auto-segmentation directly to treatment-related toxicity or patient outcomes. Future research should prioritize multi-institutional validation, clinically meaningful evaluation, improved segmentation of small glands, and human-in-the-loop integration into adaptive radiotherapy. Overall, auto-segmentation represents a promising tool for salivary gland delineation, but its clinical implementation should be supported by evidence extending beyond geometric accuracy alone.