3rd ESTRO Forum 2015 S509 segmentation method that uses Dirichlet process priors to classify the images (code dpmixsim implemented in R, R Foundation for Statistical Computing, Vienna, Austria). The DP mixture model is a Bayesian method based on Markov chain Monte Carlo simulations for exploring mixture models with an unknown number of components, not specified in advance. It depends on 4 parameters (M for precision of the DP; a and b related to prior distribution; minvar for the minimum value

more »

... dmissible for a cluster variance). We repeated contour generation on the bigger sphere of the IEC phantom by setting the DP parameters to different values and we performed an anova test to verify their ability to reproduce the true radius of the sphere. Once the optimal parameter set was found, we applied the proposed algorithm to the whole dataset. Volumes obtained by the DP algorithm were compared to the true values of the IEC spheres and of the digital phantom, and to the volumes retrospectively segmented by an experienced radiation oncologist for the 20 clinical cases. Results: The only parameter that influenced lesion segmentation was minvar, that depended almost linearly on the standard deviation of the voxel values in the region of interest. The parameters used for the analysis were: M=1; a=1; b=0; minvar optimized based on σ ROI . The agreement between the reference volumes and the result of the segmentation was within 5% for the spherical phantom and within 10% for the digital phantom and the clinical cases. Conclusions: The described procedure allowed a robust, automatic segmentation of PET volumes to be performed that accurately described reference values. This might form the basis for clinical implementation of the algorithm. Purpose/Objective: Recently there has been considerable clinical interest in the use of MRI to delineate gross target volumes (GTV) and organs at risk (OAR) for head and neck squamous cell carcinoma (HNSCC). A gold standard (GS) for delineation is considered to be delineation using a dedicated treatment position MRI (MRI-RT) rigidly registered to the planning CT scan, although the MRI-RT is not widely available in many centres. This study aimed to assess whether deformable image registration (DIR) of a diagnostic position MRI (MRI-D) to the planning CT scan is an adequate surrogate for the GS. Materials and Methods: A prospective pilot imaging study was performed with 3 HNSCC patients (oropharynx, larynx and hypopharynx cancers) who underwent contrast enhanced CT and T1 weighted MRI both in (MRI-RT) and out of (MRI-D) an immobilisation mask. A Radiation Oncologist delineated GTV and OARs (see Table 1 ) on CT, MRI-RT and MRI-D independently each on 3 separate occasions. Contour comparison (parameters shown in Table 1 ) was performed with ImSimQA v3.1.5 (OSL, Shrewsbury UK) to assess intraobserver variability of contouring on each imaging modality. Consensus contours from the 3 delineations were produced using a form of majority vote. GS structures were defined as consensus contours from MRI-I transposed to CT using rigid registration (simulating radiotherapy departments with a dedicated MRI scanner). The GS was compared to contours produced by 2 other methods: MRI-D transposed to CT with DIR (simulating radiotherapy departments without a dedicated MRI scanner); and CT alone delineations (simulating radiotherapy departments with no MRI access). All registrations were performed using Mirada RTx v1.4 (Mirada Medical, Oxford UK). Results: Contouring on MRI (MRI-RT or MRI-D) reduced intraobserver variability compared to CT (Dice similarity coefficient (DSC) range 0.82-0.99 for MRI and 0.76-0.85 for CT, Table 1 ). Contouring on CT alone is less accurate than the GS, with particularly limited accuracy compared with the GS for GTV and cord. Delineating on MRI-D and using DIR to transpose contours onto the planning CT appears superior to delineation on CT alone for GTV (DSC 0.78 versus 0.5); there was no clear benefit compared with CT alone for OAR. Conclusions: Reproducibility of contouring with MRI was found to be better than with CT indicating that the addition of MRI to the workflow for HNSCC patients is preferable. Contouring on MRI-RT was more accurate than contouring on CT or MRI-D and therefore a preferable workflow. However,