Optimizing Atlas Counts for MRI-Guided Atlas-Based Autosegmentation of Swallowing Muscles in Head and Neck Radiotherapy

PurposeRadiotherapy-induced dysphagia can significantly impair head and neck (H&N) cancer patients quality of life. Despite the dose-dependent relationship between radiotherapy and dysphagia, swallowing structures are not routinely contoured due to time and labor demands. We evaluated atlas-based autosegmentation (ABAS) on MRI, identifying the optimal number of atlases required to efficiently and accurately delineate swallowing structures. MethodsThis study included pre-radiotherapy simulation T2-weighted MRIs from 60 H&N cancer patients enrolled in an IRB-approved observational trial. Scans were acquired on a 1.5T Siemens Aera scanner with H&N immobilization. Swallowing structures, including epiglottis, constrictors, digastric muscles, genioglossus, and others, were manually contoured for 25 atlas patients and 35 test patients. GTV-involved structures were excluded. ABAS was performed with increasing numbers of atlases (1-25) using a random-forest algorithm (ABAS-ADMIRE; Elekta) to determine the optimal atlas count. To mitigate variability from atlas selection, bootstrap resampling was implemented. Dice similarity coefficient (DSC), surface DSC (SDSC), average surface distance (ASD), and 95% Hausdorff distance (HD95) were calculated for each structure. Median computation times were calculated for each atlas count. Hsus MCB analysis identified the minimum atlas number statistically equivalent to the best-performing atlas range. ResultsAcross all structures and metrics, Hsus analysis demonstrated that 2-4 atlases performed similarly to the best-performing atlas count. All structures except constrictors achieved median DSC>0.75 with [≥]2 atlases. Computation times increased linearly with atlas count (range: [~]22-950 seconds for 1-25 atlases). These findings highlight that smaller atlas counts achieve comparable accuracy while significantly improving time efficiency. ConclusionAtlas-based autosegmentation is useful for delineating swallowing muscles in radiotherapy, especially with limited available contoured datasets. Utilizing 2-4 atlases achieves similar geometric accuracy to larger atlas counts, significantly reducing computational time without compromising clinical quality. This balance between efficiency and accuracy supports integration into workflows for better dysphagia prediction and treatment planning.