ESTRO 2020 Abstract book

S284 ESTRO 2020

OC-0472 A potential multi-model AI framework for MRI motion correction and fast image acquisition in RT M. Lempart 1 , C. Jamtheim Gustafsson 1,2 , P. Brynolfsson 2 , M. Nilsson 3 , A. Gunnlaugsson 1 , L. E. Olsson 1,2 1 Skåne University Hospital, Hematology- Oncology and Radiation Physics, Lund, Sweden ; 2 Lund University, Translational Sciences- Medical Radiation Physics, Malmö, Sweden ; 3 Lund University, Centre for Mathematical Sciences, Lund, Sweden Purpose or Objective Magnetic resonance imaging (MRI) is known to add valuable information to external radiation therapy (ERT) treatments, such as superior soft tissue contrast, advantageous for structure delineation. Nevertheless, image acquisition takes a certain time, and in some cases, images can be affected by artefacts due to periodic motion, for example from breathing. In this study, we propose an artificial intelligence (AI) based multi-model deep learning framework for prostate cancer patients, with the potential of accelerated image acquisition using undersampled data and reducing periodic motion artefacts. Material and Methods Motion artefacts and undersampled image acquisition were simulated by augmenting 14390 MRI images of 426 prostate patients in a hybrid k-space, obtained by applying 2D Fourier transformations. To simulate breathing motion artefacts, a signal phase gradient was applied to the k- space data, resulting in a 20mm movement (anterior- posterior) of the object. Lines from the augmented k- space randomly replaced the corresponding lines in the k- space of the original image. Undersampled images were simulated by periodically removing every second k-space line, resulting in folding. For the suggested framework, a deep neural network-based classifier as well as two convolutional autoencoders (CAE) with skip connections were trained, with the dataset split into a training (70%), validation (20%) and test (10%) set. The CAEs were trained using a mean squared error (MSE) loss, where one uses motion artefact images as its input, and the other undersampled images. The original images were used as labels (ground truth). In order to distinguish between artefact free, undersampled or images with motion artefacts, the classifier network was trained using k-fold cross validation (k = 5). Depending on the result, input images were sent through either one of the two CAEs, or no changes were needed. Training of the CAEs was performed using categorical cross-entropy loss, with Softmax activation in the last layer. The performance of the CAEs was measured using the structural similarity index (SSIM) as well as peak signal to noise ratio (PSNR) for undersampled images, while the classifier performance was judged by its accuracy.

Figure 2 shows the peak dose distribution in the head phantom after a full-arc beam rotation. The dose ratio between the target (center of the phantom) and the bone (entrance dose) improved from 0.05 for a single field to 2.15 for a full-arc rotation.

Conclusion We found parameters for a compact, divergent MRT source leading to a PVDR above 20 in all depths, which is comparable to the same field size at the ESRF [2]. For constructing the source, these parameters need to be balanced with manufacturing requirements and clinical needs such as reduced peak entrance doses. References [1] Bartzsch S, Oelfke U. Phys Med Biol 2017. 62(22): 8600– 15. [2] Martínez-Rovira I, Sempau J, Prezado Y. Med Phys 2011. 39(1): 119–31.

Results