ESTRO 37 Abstract book

S98

ESTRO 37

Conclusion Gated delivery during repeated breath-holds under real- time MR-guidance with video-feedback to patients resulted in approximately 95% geometric coverage of the GTV by the PTV. High duty-cycle efficiencies were realized using this approach. OC-0186 Real-time long-term multi-object tracking on cineMR using a tracking-learning-detection framework J. Dhont 1 , D. Cusumano 2 , L. Boldrini 3 , G. Chiloiro 3 , L. Azario 2 , F. Cellini 3 , M. De Spirito 2 , L. Omelina 4 , J. Vandemeulebroucke 4 , D. Verellen 5 , V. Valentini 3 1 Universitair Ziekenhuis Brussel, Radiotherapy Medical Physics, Brussels, Belgium 2 Fondazione Policlinico Universitario A.Gemelli, UOC Fisica Sanitaria, Rome, Italy 3 Fondazione Policlinico Universitario A.Gemelli, Radioterapia Oncologica- Gemelli-ART, Rome, Italy 4 Vrije Universiteit Brussel, VUB-Department of Electronics and Informatics ETRO, Brussels, Belgium 5 GZA Ziekenhuizen - Iridium Kankernetwerk, Radiotherapy Medical Physics, Antwerp, Belgium Purpose or Objective Cine-MR imaging during radiotherapy allows accurate target monitoring. However, current tumor tracking algorithms suffer from drift and fail if the target dis- and re-appears. Furthermore, multi-object (MO) tracking is currently not clinically available. MO tracking would allow gating based on target-in-field but also OAR-not-in- field. This study adapts, applies and evaluates a tracking- learning-detection (TLD) framework which allows real- time and long-term MO tracking in cine-MR. It allows dis- and re-appearance of objects that move perpendicular to the 2D MR slice. The TLD framework was proposed elsewhere [1], but to the knowledge of the authors the application has been limited to non-medical images such as high frame-rate action videos. Material and Methods The TLD framework consists of three building blocks; a tracker, detection algorithm and a learning framework allowing the detection algorithm to learn in real-time, not requiring off-line pretreatment learning. The tracking algorithm applied in this study is based on median flow. The object detector uses a scanning-window grid and cascaded classifier. Each patch is re-sampled to a normalized resolution (25x25 pixels). The cascaded classifier consists of patch variance analysis, followed by an ensemble classifier performing independent pixel comparisons and finally a nearest neighbor classifier. The detection algorithm is trained based on the first cine-MR frame on which the objects to be tracked are indicated by the user. On each of the following frames, both the tracker and the object detector individually locate the objects. Exploiting the fact that the objects follow a smooth trajectory with limited frame-to-frame motion, and that the objects can only be in one location, detector errors (false positives and - negatives) can be estimated. The labels of wrong detections are corrected and fed back to the detector to train, improving its accuracy. The final object location is determined based on the most confident result from either tracker or detector. If neither give a location, the object is not visible. The TLD framework was evaluated based on the center- of-mass of 10 objects in sagittal 2D cine-MR (0.35 T, 4 Hz, 3 x 3 mm) from 5 patients treated on Viewray MRIdian, see Fig 1. The objects ranged from unique and stable in appearance (obj. 1,3,6,8) to highly deforming (obj. 4,7,9,10), or were very similar to other objects in the image (obj. 2,5,9). Ground-truth was established through contouring by an expert. Intra-observer variability was evaluated by contouring twice on different days. Results

Object tracking accuracy can be found in Table 1, together with intra-observer variability. Non-optimized, the TLD framework could be applied simultaneously to three objects on a general desktop PC in real-time.

Conclusion The TLD framework seems promising for accurate (subpixel < 3 mm) and long-term MO tracking on clinical cine-MR images. A tracking component based on deformable registration is the next step to further improve the accuracy.

OC-0187 How the sampling strategy of 2D MRI affects imaging latencies in real-time MR-guided radiotherapy P. Borman 1 , H.N. Tijssen 1 , C. Bos 2 , C.T.W. Moonen 2 , B.W. Raaymakers 1 , M. Glitzner 1 1 UMC Utrecht, Radiotherapy, Utrecht, The Netherlands 2 UMC Utrecht, Imaging Sciences Institute, Utrecht, The Netherlands Purpose or Objective The ultimate goal of MR guided radiotherapy is to adapt the treatment to anatomic changes due to e.g. respiratory motion, using a continuous stream of MR images. To react as quickly as possible, it is vital to minimize the latency between the occurrence of anatomic change and its appearance on the image. Since MR imaging is relatively slow with respect to other imaging modalities, it is necessary to use its acquisition flexibility to optimize latency and speed. In this work we present a systematic analysis of the latencies inherent to real-time MRI and show how the choice of sampling pattern and use of Partial Fourier (PF), influences the apparent imaging latency, i.e. the difference between actual and observed position after reconstruction. Material and Methods Partial Fourier is a common acceleration technique in MR, where only part k-space is acquired and missing parts are synthesized in reconstruction. As such it was used to accelerate the acquisition (full matrix: 256x256). In addition, ‘linear’, ‘reverse linear’ and ‘high-low’ sampling patterns were used to investigate the effects on apparent image latencies (fig. 1). The experiments were performed on the clinical prototype of the Elekta Unity system (1.5T high-field MR). Motion was generated using a 4D motion phantom (QUASAR MRI4D, modusQA). The phantom was set to perform a 1D sinusoidal trajectory of