Automatic Segmentation To Define Organs At Risk (Oars) For Function Sparing Head And Neck Imrt

The introduction of function sparing and adaptive IMRT for head and neck cancer has increased the number and frequency of delineation of OARs. We investigated whether automatic segmentation with Smart Probabilistic Image Contouring Engine (SPICE) may reduce contouring time, decrease variability and maintain accuracy. Ten radiation therapy planning CT scans were selected where the OARs of interest were not distorted by tumor or artefact. Five clinicians (4 Attending Physicians and 1 Fellow) recorded for each scan the time to delineate the parotid and submandibular glands, larynx, pharyngeal constrictor muscles and cochleae according to published guidelines. SPICE was then used to define these structures. Each clinician determined by visual inspection the acceptability of SPICE defined volumes and the total time to modify contours for each patient. The Simultaneous Truth and Performance Level Estimation (STAPLE) algorithm created a reference standard from all clinician contours. Clinician and SPICE defined contours were compared to STAPLE by three matrices: conformity index (CI, ratio of volumes, target = 1.0), Dice similarity coefficient (DSC, 2x intersection volume/sum of total volumes, target = 1.0, acceptable>0.7) and distance to agreement (DTA), the per voxel shortest distance from the surface of one structure to another (maximum and mean DTA, target = 0.0mm). The proportions of SPICE defined contours deemed not to require alteration were: parotid glands (17%), submandibular glands (41%), larynx (8%), pharyngeal constrictor muscles (4%) and cochleae (28%). Comparisons between the clinicians’ contours and STAPLE (the reference standard) versus SPICE and STAPLE are shown in the Table. For the pharyngeal constrictor muscles and larynx, SPICE compared unfavorably for all matrices. For the parotid glands, only the DSC was significantly different and this remained acceptable for SPICE (median, 0.79). The DSC and DTA for SPICE defined submandibular glands were inferior. This was partly due to low inter-observer variability in clinician contours and the SPICE DSC was acceptable (median, 0.80). The SPICE defined cochleae had a low DSC (median, 0.30). The overall respective mean times to manually define and modify the automatic contours for each patient were 14.0 minutes and 16.2 minutes (p<0.01). The accuracy of automatic segmentation requires improvement to enable increased efficiency in the definition of OARs for head and neck IMRT.Scientific Abstract 3693; TableComparison between all clinicians and SPICE defined OARs and the reference standard (STAPLE)OAR (N clinicians, N SPICE)DSC (median, IQR)Maximum DTA / mm (median, IQR)Mean DTA / mm (median, IQR)CliniciansSPICECliniciansSPICECliniciansSPICEParotid glands (N = 100, N = 20)0.90, 0.88-0.920.79, 0.74-0.83***7.8, 5.2-11.114.8, 11.6-19.10.86, 0.34-1.31.5, 0.93-2.0Submandibular glands (N = 100, N = 20)0.91, 0.88-0.930.80, 0.70-0.85***3.8, 2.9-5.15.7, 4.7-8.2***0.23, 0.11-0.500.61, 0.41-1.0***Larynx (N = 50, N = 10)0.85, 0.81-0.870.57, 0.54-0.62***6.4, 5.5-7.822.4, 15.8-25.4***0.45, 0.25-0.825.2, 2.4-6.3***Pharyngeal constrictor muscles (N = 50, N = 10)0.76, 0.71-0.790.53, 0.39-0.57***6.0, 5.1-7.512.9, 8.7-15.9***0.76, 0.53-0.921.2, 0.93-1.9*Cochleae (N = 100, N = 20)0.63, 0.48-0.810.30, 0.13-0.49***3.0, 1.3-3.32.6, 1.8-5.01.1, 0.18-1.80.89, 0.65-3.3Abbreviations: OAR, organ at risk; N = total number of OARs; CI, conformity index; DSC, Dice similarity coefficient; DTA, distance to agreement; IQR, interquartile range. Comparison: All clinicians versus STAPLE compared with SPICE versus STAPLE; *P