Auto-segmentation of low-risk clinical target volume for head and neck radiation therapy

Purpose: To investigate atlas-based auto-segmentation methods to improve the quality of the delineation of low-risk clinical target volumes (CTVs) of unilateral tonsil cancers.

Method and materials: Sixteen patients received intensity modulated radiation therapy for left tonsil tumors. These patients were treated by a total of 8 oncologists, who delineated all contours manually on the planning CT image. We chose 6 of the patients as atlas cases and used atlas-based auto-segmentation to map each the atlas CTV to the other 10 patients (test patients). For each test patient, the final contour was produced by combining the 6 individual segmentations from the atlases using the simultaneous truth and performance level estimation algorithm. In addition, for each test patient, we identified a single atlas that produced deformed contours best matching the physician's manual contours. The auto-segmented contours were compared with the physician's manual contours using the slice-wise Hausdorff distance (HD), the slice-wise Dice similarity coefficient (DSC), and a total volume overlap index.

Results: No single atlas consistently produced good results for all 10 test cases. The multiatlas segmentation achieved a good agreement between auto-segmented contours and manual contours, with a median slice-wise HD of 7.4 ± 1.0 mm, median slice-wise DSC of 80.2% ± 5.9%, and total volume overlap of 77.8% ± 3.3% over the 10 test cases. For radiation oncologists who contoured both the test case and one of the atlas cases, the best atlas for a test case had almost always been contoured by the oncologist who had contoured that test case, indicating that individual physician's practice dominated in target delineation and was an important factor in optimal atlas selection.

Conclusions: Multiatlas segmentation may improve the quality of CTV delineation in clinical practice for unilateral tonsil cancers. We also showed that individual physician's practice was an important factor in selecting the optimal atlas for atlas-based auto-segmentation.