Adam Szmul & Yuliang Huang - CMIC/WEISS Joint Seminar Series

Speaker:  Adam Szmul

Title:  Towards Adaptive Radiotherapy by CT Synthesis and Structure Segmentation from Daily Cone Beam CTs

Abstract: 
Adaptive radiotherapy (ART) has the potential to improve the outcomes of the therapy both in terms of better tumour control and reduced healthy tissue toxicity. This requires images of comparable quality to the planning Computed Tomography (CT) for anatomical structures segmentation and dose recalculation. Daily acquired Cone Beam Computed Tomography (CBCT) images are of too low quality to be directly used in that application.
We developed a novel approach for synthesising CTs from daily CBCT, using cycle consistent Generative Adversarial Networks (cycleGANs). That method was tailored for paediatric abdominal patients, which is particularly challenging due to the inter-fractional variability in bowel filling. We introduced global residuals only learning to the networks, redefining the learning task as removing unwanted artefacts from the source CBCTs. We modified the cycleGAN loss function to explicitly promote structural consistency between source and synthetic images, by adding a term based on normalised cross correlation. Finally, to compensate for the considerable anatomical variability in the paediatric population and address the difficulties in collecting large datasets, we applied a smart 2D slice selection based on the common (abdominal) field-of-view across the dataset, acting as a weakly paired data approach.  
We present how we are investigating a joint CT synthesis and structure segmentation from daily CBCT. That process will be conditioned by planning CT and corresponding structures delineated from it prior to the treatment.

Speaker:  Yuliang Huang

Title:  Motion modelling for dynamic CBCT reconstruction

Abstract: 
Cone-beam CT (CBCT) is widely used in image guided radiotherapy, but motion due to breathing can blur the image. Similar to 4DCT, 4D CBCT can reduce motion blur but 4D CBCT acquisitions take 2~4 times longer than 3D CBCT and often suffer from phase sorting artefact. This study aims to obtain motion models and motion-free images simultaneously from unsorted 3D CBCT projection data, using a general motion modelling framework previously proposed by our group, which was for the first time applied to CBCT equivalent to a 1min acquisition.

Chair: Jamie McClelland