Program Information
Gaussian Weighted Multi-Atlas Based Segmentation for Head and Neck Radiotherapy Planning
M Peroni1*, GC Sharp2, P Golland3, G Baroni1,4, (1) Department of Bioengineering, Politecnico di Milano, Milano, Italy, (2) Department of Radiation Oncology, Massachusetts General Hospital, Boston, MA, (3) Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, MA, (4) Bioengineering Unit, Centro Nazionale di Adroterapia Oncologica, Pavia, Italy
WE-E-213CD-2 Wednesday 2:00:00 PM - 3:50:00 PM Room: 213CDPurpose: To develop a multi-atlas segmentation strategy for IMRT head and neck therapy planning.
Methods: The method was tested on thirty-one head and neck simulation CTs, without demographic or pathology pre-clustering. We compare Fixed Number (FN) and Thresholding (TH) selection (based on normalized mutual information ranking) of the atlases to be included for current patient segmentation. Next step is a pairwise demons Deformable Registration (DR) onto current patient CT. DR was extended to automatically compensate for patient different field of view. Propagated labels are combined according to a Gaussian Weighted (GW) fusion rule, adapted to poor soft tissues contrast. Agreement with manual segmentation was quantified in terms of Dice Similarity Coefficient (DSC). Selection methods, number of atlases used, as well as GW, average and majority voting fusion were discriminated by means of Friedman Test (a=5%). Experimental tuning of the algorithm parameters was performed on five patients, deriving an optimal configuration for each structure.
Results: DSC reduction was not significant when ten or more atlases are selected, whereas DSC for single most similar atlas selection is 10% lower in median. DSC of FN selection rule were significantly higher for most structures. Tubular structures may benefit from computing average contour rather than looking at the singular voxel contribution, whereas the best performing strategy for all other structures was GW. When half database is selected, final median DSC were 0.86, 0.80, 0.51, 0.81, 0.69 and 0.79 for mandible, spine, optical nerves, eyes, parotids and brainstem respectively.
Conclusion: We developed an efficient algorithm for multi-atlas based segmentation of planning CT volumes, based on DR and GW. FN selection of database atlases is foreseen to increase computational efficiency. The absence of clinical pre-clustering and specific imaging protocol on database subjects makes the results closer to real clinical application.
Contact Email