Deep-learning-assisted brain contouring and serial neurocognitive monitoring in head and neck cancer survivorship: a prospective study.

Patients treated for head and neck cancer (HNC) receive incidental, non-target radiation to intracranial structures. This prospective study investigated the relationship between regional brain radiation dosimetry and neurocognitive performance during the first year of survivorship. Patients with newly diagnosed HNC completed a comprehensive neuropsychological battery at baseline-prior to or within the first week of radiation therapy, or before induction chemotherapy-(n = 47), and again at 6 months (n = 30) and 12 months (n = 24) after initiating radiation therapy. Cognitive domains assessed included language, learning and memory, and attention, processing speed, and executive function (APE). Deep-learning-assisted contouring was applied to estimate mean radiation doses (Dmean) to eight brain regions. Associations between regional dosimetry and cognitive change were evaluated using Pearson correlation coefficients with 95% confidence intervals (CIs). The highest median Dmean were observed in the cerebellum (430.7 cGy; IQR: 302.3-646.2) and temporal lobes (108.5 cGy; IQR: 72.4-161.2), while the parietal lobe received the lowest (27.5 cGy; IQR: 18.9-39.6). Neurocognitive scores demonstrated modest improvement across domains at both follow-up time points. No significant correlations were identified between region-specific doses and cognitive performance, with most estimates between -0.1 and 0.1. Incidental brain radiation during HNC treatment was not associated with early neurocognitive decline within the first year after diagnosis. This prospective study illustrates the utility of integrating advanced dosimetry with longitudinal cognitive assessments and lays the groundwork for future larger-scale investigations.