ESTRO 2021 Abstract Book

S660

ESTRO 2021

Inter-observer variation in contouring tumour volumes can be significant and often depends on the experience of the operator. Other modalities, such as PET, are used for contouring in addition to CT for head and neck cancer (HNC). In recent years, convolutional neural networks (CNNs) have been developed for contouring and have shown good agreement with manual contouring, with a reduction in observer variability. However, combining different imaging modalities raises questions about the contribution of each modality to the auto- contouring process. The aim of this study was to understand the contribution of each imaging modality to contour accuracy and to assess the effect of misalignment between PET and CT modalities (in the case that scans were acquired separately). Materials and Methods 201 PET-CT cases from the MICCAI 2020 challenge were used. All cases had oropharynx tumours with Gross Target Volume (GTV) defined by an expert. To understand the impact of rigid alignment error on contour accuracy, multiple CNNs were trained with varying degrees of misalignment introduced between the PET and CT. Two approaches were taken; one set of models was trained keeping the GTV contour aligned to the CT, while a second set was trained with the GTV aligned to the PET. Each model within a set used varying degrees of misalignment between the images. A similar approach was taken with the test data, whereby the GTV contour was correctly aligned with either the CT (denoted ctGTV) or PET (denoted petGTV). Dice similarity coefficient (DSC) was calculated to compare the overlap of the CNN-predicted GTV (denoted pGTV) with the ctGTV and petGTV. The agreement between the ctGTV and petGTV was also calculated as a reference to indicate the degree of misalignment. Results As illustrated in Figure 1, the performance of the automated contour delineation was better when the model was trained with petGTV compared to ctGTV. Training using the petGTV gave a consistent performance regardless of the degree of shift applied either at training or test. Moreover, the DSC between ctGTV and pGTV is similar to the one between that of ctGTV and petGTV. If the shift between the CT and PET at test time is greater than the shift between the modalities in training, then the model trained using the ctGTV fails.

Conclusion Training a CNN using the petGTV gave superior results compared to one trained using ctGTV. The petGTV- trained CNN was more consistent regardless of the degree of shift and so less sensitive to alignment errors between CT and PET. This highlights that the CNN-generated contour predominantly follows petGTV and that hybrid CT-PET CNNs pay more attention to PET than CT images. Clinically, any misalignment between PET and CT will result in biased contours in the planning CT, which is a problem that needs further investigation. Acknowledgement: This work was supported by the Cancer Research UK Radnet grant CRUK A28736 PD-0826 Super-resolution ultrasound and MRI imaging for monitoring breast tumour response to radiotherapy M. Morris 1,2 , M. Toulemonde 1 , V. Sinnett 3 , S. Allen 4 , K. Downey 4 , N. Tunariu 2,4 , C. Lucy 2,5 , L. Gothard 2,5 , G. Hopkinson 4 , E. Scurr 4 , E. Harris 2,6 , M. Tang 1 , M. Blackledge 2 , N. Somaiah 2,5 1 Imperial College London, Bioengineering, London, United Kingdom; 2 The Institute of Cancer Research, Radiotherapy and Imaging, London, United Kingdom; 3 The Royal Marsden NHS Foundation Trust, Radiology , London, United Kingdom; 4 The Royal Marsden NHS Foundation Trust, Radiology, London, United Kingdom; 5 The Royal Marsden NHS Foundation Trust, Breast Unit, London, United Kingdom; 6 The Royal Marsden NHS Foundation Trust, Physics, London, United Kingdom Purpose or Objective There are currently no specific protocols to stratify breast cancer patients based on tumour sensitivity to radiotherapy (RT). Tumours are associated with high cellularity, vascular permeability and chaotic microvasculature. Our hypothesis is that a novel imaging technique, Super-Resolution UltraSound (SRUS) [1] combined with biomarkers from dynamic-contrast enhanced and diffusion-weighted MRI (DCE/DW-MRI) will enable unprecedented quantification of tumour microvessel structure and dynamics. This will allow