ESTRO 2020 Abstract book

S1017 ESTRO 2020

PO-1740 Quantitative evaluation of ultra-low dose paediatric cone beam CT for image-guided radiotherapy A. Bryce-Atkinson 1 , R. De Jong 2 , A. Bel 2 , M.C. Aznar 1 , G. Whitfield 3 , M. Van Herk 1 1 The University of Manchester, Division of Cancer Sciences- School of Medical Sciences- Faculty of Biology Medicine and Health, Manchester, United Kingdom ; 2 Amsterdam University Medical Center, Department of Radiation Oncology, Amsterdam, The Netherlands ; 3 Royal Manchester Children's Hospital, The Children's Brain Tumour Research Network- The University of Manchester, Manchester, United Kingdom Purpose or Objective Daily cone beam CT (CBCT) for image-guided radiotherapy (IGRT) is more reluctantly applied to children due to concerns over imaging dose. Children are more radiosensitive than adults, and coupled with their longer life expectancy, are more at risk to suffer from radiation induced late effects and secondary malignancies. Limiting CBCT dose is paramount such that children can receive the benefit of accurate IGRT with limited dose burden. Simulating low dose CBCT allows assessment of imaging protocols without additional dose to the patient. This work simulates ultra-low dose CBCT and evaluates its use for paediatric IGRT by quantitative assessment of image registration accuracy and visual assessment of image quality in an observer study. Material and Methods Ultra-low dose CBCT was simulated for 20 paediatric patients (aged 1-16 years, median 7 years) by adding signal-dependant uncorrelated Gaussian noise to projection images prior to reconstruction. The method was validated by visual assessment and comparison of contrast- noise-ratio (CNR) and signal-noise-ratio (SNR) in a Catphan phantom. Scans were simulated at a range of ultra-low doses (100 kVp, half rotation, 0.10mAs - 0.025mAs per projection). Image quality was assessed by four experienced RTTs acting as observers, assessing registration accuracy on bone and soft tissue and by Likert scale grading of image features, comparing results to the current clinical low-dose protocol (100kVp, half rotation, 0.16mAs) used in our institution. Results Simulated and acquired phantom scans were in excellent agreement, with SNR and CNR agreeing within 6% for all exposures. For patient scans, bony anatomy registration discrepancies compared to the current protocol all fell within 2mm (translation) or 1 degrees (rotation) with inter-observer variation comparable to the existing protocol (figure 1). Soft tissue registration showed large discrepancies in both ultra-low and current protocols. Registration discrepancies over 3mm were deemed as failed registrations. In the full dose scans (0.16mAs), there was 20% incidence of failure in soft tissue registration, rising inconsistently in the lower dose scans: to 58%(0.10mAs), 37.5%(0.08mAs), 39%(0.04mAs) and 42%(0.025mAs). Figure 2 shows the visual image quality at the extreme of simulated exposures. In the Likert grading, bony anatomy visualisation and registration reached over 75% acceptability (rated ‘well’ or ‘very well’) for all doses, however soft tissue visualisation did not reach this threshold at any dose, reflecting the quantitative registration results.

position into one of the CBCT scans without metal artefacts. This scan was then forward projected, beam hardening and scatter were added, and the scan was back- projected again, thus creating an artificial metal artefact. A ground truth image was created using the same process, but without adding the beam hardening and scatter in projection space. An image-based 2D U-net network structure was used, taking the slice with the metal artefact as well as the two neighbouring slices as input. The network was trained on 3000 samples, to output the difference between the corrupted input slice and the ground truth. Thus, when evaluating the network, the corrected image is the input slice minus the CNN output. The corrected image was post processed with a frequency-split edge-preserving method [Med. Phys. 39(4):1904-1916, 2012], enhancing the detail information from the input image in the output image. The network was evaluated on its ability to correctly identify and remove the streaks and as compared to a current state-of-the-art frequency-split normalised metal artefact reduction (FSNMAR) algorithm [Med. Phys. 37(10):5482-5493, 2010]. The two methods were compared visually using a clinical CBCT scan with metal artefacts. Results An example of a clinical CBCT with real metal artefacts and the networks ability to remove it is illustrated in Figure 1.

Figure 1 also show the FSNMAR correction for comparison. In some areas the CNN metal artefact reduction (CNNMAR) algorithm provides better recovery of soft tissue structures and bone edges, while in other areas the FSNMAR is similarly superior. The effect of the frequency-split on the CNNMAR algorithm can be seen in Figure 2, restoring edge information in the immediate vicinity of the metal implant.

Conclusion One correction method does not seem to be strictly better than the other, however in terms of computational time, the CNNMAR is faster. The network does not recreate anatomy obscured by the artefact streaks correctly, and this could probably be improved by training on data from more patients to increase the generalisability of the network, potentially becoming both strictly better and faster than FSNMAR.