ESTRO 2020 Abstract Book

S183 ESTRO 2020

efficiency for the junior RO group, which could facilitate implementation of institutional contouring standards and improvement of practice consistency. OC-0344 Automatic contouring of diffusion-weighted MRI from an MR-Linac using a convolutional neural network O. Gurney-Champion 1 , J. Kieselmann 1 , W. Kee 2 , B. Ng- Cheng-Hin 3 , K. Newbold 2 , K. Harrington 3 , U. Oelfke 1 1 The Institute of Cancer Research and The Royal Marsden NHS Foundation Trust, The Joint Department of Physics, London, United Kingdom ; 2 The Royal Marsden NHS Foundation Trust, Clinical Oncology, London, United Kingdom ; 3 The Institute of Cancer Research and The Royal Marsden NHS Foundation Trust, Targeted Therapy team, London, United Kingdom Purpose or Objective The MR-Linac allows for easy daily acquisition of diffusion- weighted imaging (DWI). Retrieving quantitative MRI information (the apparent diffusion coefficient; ADC) from DWI involves contouring regions of interest. This contouring is time-consuming, particularly with daily repeated DWI, and is prone to manual errors. Therefore, we trained a deep convolutional neural network that automatically and systematically contoured metastasized lymph nodes on DWI of head and neck cancer (HNC) patients. Material and Methods DWI images from 48 HNC patients with a total of 65 metastasized lymph nodes were obtained on a diagnostic 1.5T MR-scanner throughout radiotherapy (RT) treatment (18 with induction chemotherapy; 30 RT-only). A radiation oncologist delineated these lymph nodes on the b=50 s/mm 2 diffusion-weighted source images. To assess the inter-observer variability, a second radiation oncologist also contoured the lymph nodes for five of the images. We implemented a 3D U-net in python using Keras and Tensorflow (Dice loss function; 20% dropout; batch normalisation; 64 base features; 4 pooling layers; 22 convolutional layers; learning rate: 2e-4; automated stopping criteria with latency of 20 epochs). We added a voxel-wise bias-layer [Dalca et al. IEEE-CVPR 2018] before each convolution. We envision a workflow in which a clinician can receive contours and ADC information by clicking on the lymph node. In this workflow, a bounding box (64×64×32 voxels) is placed centred at the selected voxel and used as input for the U-net. Such clicks were simulated and training was done using the resulting bounding boxed b=50 s/mm 2 DWI image and the corresponding manual contours. The performance was evaluated using 8-fold cross- validation (over patients) and calculating the Dice similarity coefficient (DSC) and absolute change in median ADC between the manual and the learnt contours. Finally, the performance of the network was tested on an independent test dataset with 4 DWI images (3 patients) including 8 involved lymph nodes obtained with a substantially different imaging protocol on a 1.5T MR- Linac.

Results The DSC and root-mean-square of the change in median ADC were on average 0.84 and 3.0%, respectively, for the RT-only patients (Table 1). These numbers were similar for the different treatment time points (1 and 2 weeks into RT). The inter-observer variability had a DSC of 0.86 whereas the network achieved a DSC of 0.89 in the same 5 patients. The network performed poorer on the patients receiving induction chemo, potentially due to chemo affecting the clarity of the tumour borders. The network performed well on the independent MR-Linac dataset (Fig 1) with an average DSC of 0.81 and a change in median ADC of 2.0%. This demonstrates its potential to generalise between systems. Conclusion We enabled automatic and accurate contouring of metastasized lymph nodes in the head and neck region on diagnostic and MR-Linac DWI images. This will help enable a feasible MR-Linac workflow with daily DWI imaging.