ESTRO 37 Abstract book

S1182

ESTRO 37

Research, Odense C - DK-5000, Denmark 3 Princess Margaret Cancer Centre, Department of Radiation Oncology, Toronto ON- M5G 2M9, Canada Purpose or Objective 4D-Cone Beam CT (4D-CBCT) images used for image guided radiotherapy (IGRT) are visually hampered by streaking artifacts. A time efficient correction for small field of view (SFOV) 4D-CBCT scans is investigated. Material and Methods This study is based on SFOV 4D-CBCT and 4D-CT scans of 20 non-small-cell lung cancer patients subjected to curatively intended radiotherapy initiated between March 2012 and August 2014. CBCT projections were FDK- reconstructed and forward projected by software in the RTK-package (www.openrtk.org). The applied streak reduction approach resembles elements of the McKinnon- Bates algorithm [IEEE Trans Biomed Eng., 28, 123, 1981]. Initially all CBCT projections were 3D reconstructed and re-projected onto the original projection angles. Re- projections were reconstructed as 3D (re3D-CBCT) and as 4D (re4D-CBCT) by phase binning extracted from the original projections. As the volumes reconstructed from re-projections were motionless, streaks occurring due to undersampling of projections were estimated from their difference (Streaks = re4D-CBCT - re3D-CBCT). Corrected 4D-CBCTs were made by subtracting the streaks estimate from a standard 4D-CBCT reconstruction (Corrected4D- CBCT = 4D-CBCT – Streaks). Total calculation time for a 10 phase reconstruction was below 5 minutes. The expiration phase was used for quantitative analyses. Streak-reduction was measured as the sum of gradient’s magnitudes in the corrected image divided by the sum of gradient’s magnitudes in the uncorrected image, with less than unity indicating a streak reduction. Potential changes in the Hounsfield units (HU) as result of the streak reduction were studied by mapping CBCT images onto CT by deformable image registration. For each CBCT image the average HU error between CT and CBCT was calculated. CBCT values were also studied by dividing each CBCT image into bins of 40 HUs based on the CT values in order to evaluate overall HU changes. Results Visual improvements were clearly observed when applying streak correction for the 4D-CBCTs (see figure). The gradient based measure of streak-reduction also favored the corrected images, having a sample mean (standard deviation) of measured streak-reduction: 0.83 (0.053). No sign of overall degradation of HU values were observed, given that sample mean (standard deviation) of average HU error for corrected CBCT was: 181 (22.1) and for uncorrected CBCT: 193 (24.7). The median of binned CBCT values remained unchanged, and even a small drop in the uncertainties of the binned CBCT values was observed when streak correction was applied.

Conclusion SFOV 4D-CBCT images were found to improve visually when subjected to a simple streak correction. Improvements were likewise observed on quantitative measurements of gradients as well as on CBCT HU to CT HU comparisons, and did not degrade the overall HU levels of the image. With gpu-optimization the calculation time of the method can potentially be reduced to less than a minute which will make it usable for on-line IGRT. EP-2140 A Bayesian mixture-model for ion identification and filter in particle imaging C.A. Collins Fekete 1 1 National Physical Laboratory, Chemical- Medical and Environmental Science, Teddington, United Kingdom Purpose or Objective Proton imaging promises to produce accurate relative stopping power (RSP) maps crucial for particle therapy treatment planning. Yet, protons suffer from multiple Coulomb scattering that degrades the image spatial resolution. Less deflected-heavier ions have been suggested to reduce this problem. However, those ions produce secondary particles through nuclear interactions, which increase the detector noise. A Bayesian mixture- model (BMM) is developed to predict the most-likely particle related to each set of measurements produced by an event. Material and Methods In a mixture model, it is assumed that a given event can be drawn from one of N generating processes. In particle imaging, it is assumed that the detector's measurements