ESTRO 36 Abstract Book

S494 ESTRO 36 2017 _______________________________________________________________________________________________

D. Montgomery 1 , K. Cheng 1 , Y. Feng 1 , D.B. McLaren 2 , S. McLaughlin 3 , W. Nailon 1 1 Edinburgh Cancer Centre Western General Hospital, Department of Oncology Physics, Edinburgh, United Kingdom 2 Edinburgh Cancer Centre Western General Hospital, Department of Clinical Oncology, Edinburgh, United Kingdom 3 Heriot Watt University, School of Engineering and Physical Sciences, Edinburgh, United Kingdom Purpose or Objective Prostate cancer is one of the few solid organs where radiotherapy is applied to the whole organ. This is because accurately identifying the dominant cancer foci on magnetic resonance (MR) images, which can then be mapped onto computerised tomography (CT) images for radiotherapy planning, is difficult. The aim of this study was to investigate the use of three-dimensional (3D) texture analysis for automatically identifying the dominant cancer foci on MR images acquired for diagnosis and prior to the administration of androgen deprivation therapy, which may shrink the tumour foci. Material and Methods On 14 patients with confirmed prostate cancer, 3D image texture analysis was carried out on T2-weighted MR images acquired for diagnosis on a Symphony 1.5T scanner (Siemens, Erlangen, Germany). The prostate, bladder, rectum and the location of the main cancer foci were outlined on all images. In 5x5x5 pixel 3 volumes within the prostate 446 3D texture analysis features were calculated. These features were used to train an AdaBoost model, which was used to predict the class of each 5x5x5 region as either 'prostate” or 'focal lesion.” Morphological filtering was applied to each region to remove invalid elements and to clean the final outline. The Dice similarity coefficient was used to assess the agreement between the clinical and predicted contours. Results Figure 1 shows an example of a contour produced by the algorithm where the Dice similarity coefficient was 0.98. Table 1 shows the Dice coefficients calculated between the clinical contours and the contours predicted by 3D image analysis. 11 of the 14 cases had a Dice score greater than 0.65 and 8 of the 14 cases had a score greater than 0.9, indicating good agreement between the clinical and predicted contours. In 3 cases the image analysis technique failed to identify the focal lesion.

PO-0903 Patient-induced susceptibility effects simulation in magnetic resonance imaging J.A. Lundman 1 , M. Bylund 1 , A. Garpebring 1 , C. Thellenberg Karlsson 1 , T. Nyholm 1 1 Umeå University, Department of Radiation Sciences, Umeå, Sweden Purpose or Objective The role of MRI is increasing in radiotherapy. A fundamental requirement for safe use of MRI in radiotherapy is geometrical accuracy. One factor that can introduce geometrical distortion is patient-induced susceptibility effects. This work aims at developing a method for simulating these distortions. The specific goal being to objectively identify a balanced acquisition bandwidth, keeping these distortions within acceptable A simulation algorithm based on Maxwell’s equations and calculations of shift in the local B-field was implemented as a dedicated node in Medical Interactive Creative Environment (MICE), which is available as a free download. The algorithm was validated by comparison between the simulations and analytical solutions on digital phantoms. Simulations were then performed for four body regions using CT images for eight prostate cancer patients. For these patient images, CT Hounsfield units were converted into magnetic susceptibility values for the corresponding tissues, and run through the algorithm. limits for radiotherapy. Material and Methods

Figure 1 : Clinical contour in blue and predicted contour generated by 3D texture analysis shown in red on three T2-weighted MR images from the same patient (Patient 6).

Table 1 : Dice coefficient between the clinical contours and the contours predicted by image analysis. Conclusion The 3D image analysis results presented are encouraging and demonstrate the potential of this approach for automatically identifying focal disease on T2-weighted MR images. However, further investigation is required to establish why the approach fails in certain circumstances and to establish the performance of the approach on a much larger patient cohort.

Figure 1: Simulated normalized local B-field for one of the patients [ppm]. Results The digital phantom simulations showed good agreement with analytical solutions, with only small discrepancies due to pixelation of the phantoms. For a bandwidth of 440 Hz at 3 T, the calculated distortions in the patient-based