Abstract Book

S270

ESTRO 37

Results The implemented cluster analysis is able to automatically assign and subdivide dose volumes to the corresponding lesion and hence facilitates the determination of the necessary quality indices. In contrast to DBSCAN, the k-means algorithm requires the number of metastases as input. The main advantage of the k-means algorithm is that it can subdivide overlapping dose clouds from neighbouring lesions (Fig. 1). DBSCAN performs more stable for a larger number of separated lesions (Fig. 2) or in the presence of noise, as it is the case in e.g. 3D gel dosimetry measurements for plan QA purposes.

OC-0515 Cluster Analysis to Calculate Quality Indices for Different Treatments of Multiple Brain Metastases M. Reiner 1 , Y. Dinc 1 , F. Kamp 1 , K. Parodi 2 , C. Belka 1 1 University Hospital- LMU Munich, Department of Radiation Oncology, Munich, Germany 2 Faculty of Physics- Ludwig-Maximilian-Universität München, Department of Medical Physics, Munich, Germany Purpose or Objective For the treatment of multiple brain metastases there is a paradigm shift from whole brain radiotherapy to stereotactic treatment of the metastases themselves. Different approaches are available, of which the treatment of all metastases at once with one single isocentre is the most promising one. To compare and evaluate the different approaches several plan quality indices need to be determined independent of the TPS. The manual calculation of these indices is a challenging and time-consuming task, as e.g. the dose clouds of two different lesions might overlap and need to be reasonably disentangled. Material and Methods We developed a computational framework using Matlab and plastimatch to automatically determine all relevant quality indices (gradient index, conformity index, homogeneity index). A tailored version of the Paddick gradient index GI P is implemented as the volume of half the prescribed dose divided by the volume of the PTV (i.e., GI PTV = V hp /V PTV ), which better describes the dose- fall off around a lesion. To assess such GI PTV the total volume of a certain dose level needs to be divided into sub-volumes, one for each corresponding lesion. To achieve this dose clustering, we implement two different methods of Cluster Analysis, the k-means-method and DBSCAN ( d ensity b ased s patial c lustering for a pplications with n oise).

GI PTV is a useful, complementary quality index which better describes the dose fall-off around a lesion, as it is independent of the coverage of the lesion. The well- established GI P index assigns a higher gradient for the same distance between the prescribed dose value and the half-prescribed isodose level if the volume of V p is larger. This is the reason why a lower value for GI P (higher dose gradient) for one plan compared to another does not automatically mean that this plan is superior in dose fall- off compared to the volume of the PTV. This problem is overcome by the proposed GI PTV . Attention must be paid to the coverage, as a bad coverage leads to a small GI PTV . Conclusion Cluster Analysis provides a good tool for automatic calculation of quality indices for multiple lesions. The introduced GI PTV provides a more realistic description of the dose fall-off outside the PTV than the GI P, but also in this case one should take indices characterising the coverage (e.g. CI P ) into account to avoid underdosage when comparing different plans.