ESTRO 2021 Abstract Book

S1290

ESTRO 2021

4 Australian institute of Health innovation, Macquarie University, Sydney, Australia

Purpose or Objective Approximately one third of radiotherapy patients are treated with palliative intent, many of whom experience rapid anatomical changes between simulation and delivery. An online adaption system utilising AI segmentation on daily cone-beam CT (CBCT) to auto-generate a plan of the day, may account for these changes and foster sophisticated dynamic treatment. This work evaluated the accuracy of Ethos AI segmentation and subsequent auto plan dose distributions for palliative patients. Materials and Methods Ten previously-treated palliative patients (27 fractions) were replanned using the Ethos TPS. The patient datasets were duplicated and sorted into two test arms, reference (arm 1) and AI-only (arm 2), with a third arm composed of the original clinical plans (arm 3). Adaptive treatment was simulated in a test environment using patient-specific CBCTs. In the reference arm, clinicians edited all AI contours before an adaptive auto- plan was generated. In the AI-only arm, no contour edits were made. Adaptive plans generated were compared to the non-adapted (initial) plans using PTV D95% and 100% conformity index. On a subset of six fractions, the same metrics were used to evaluate the original plans with clinician-edited contours. AI-only and reference structures were compared using the DICE similarity coefficient. Results All arm 1 and 2 adaptive plans passed the primary constraint PTV D95%>95%, while three initial fractions failed. Mean PTV D95% for adaptive and initial fractions were 102.4% and 104.0%, respectively, with no statistically significant difference found (t-test P=0.29). Mean conformity indices measured 0.97 and 0.85 for adaptive and initial plans, respectively, and were found to be statistically different (t-test P=0.017). Comparison between the clinical plan and clinician contours found two out of six fractions failed the primary metric, with an average PTV D95%>95% of 107.1±38.7% (k=2). While, the mean conformity index for 100% dose was 0.95±0.03 (k=2). The D95% metric varied substantially, indicative of interfraction motion. The Dice similarity coefficients for CTV, kidneys, liver, small bowel and stomach were 0.90±0.09, 0.930±0.05, 0.98±0.02, 0.928±0.04 and 0.70±0.11, respectively, indicating accurate AI segmentation of all contours except for the stomach. Conclusion Ethos AI-driven adaption was shown to achieve better target coverage than initial plans. Furthermore, the clinician-edited targets received superior dose coverage when treated with the adapted plans, compared to the original plan. The plan analysis indicated the system could account for inter-fraction anatomy changes associated with the palliative cohort. Good agreement was shown between preliminary Ethos and clinician contours, except for the stomach, suggesting that AI-only contour edits may be acceptable for adaptive palliative patients. Future work includes online adaption in diagnostic CT-enabled palliative planning and further analysis of Ethos-generated contours. PO-1566 Evaluation of the DIRs driving a CBCT-guided online adaptive radiotherapy system A. Quinn 1 , J. Kipritidis 1 , J. Booth 1 1 Royal North Shore Hospital, Northern Sydney Cancer Centre, Sydney, Australia Purpose or Objective The Varian CBCT-guided online adaptive radiotherapy system is predicated on synthetic CT (sCT) and dose accumulation both derived from deformable image registration (DIR). In this work, we implement qualitative and quantitative tools to evaluate the sCT used at point of care and the dose accumulation DIR used for inter- fraction decisions. Materials and Methods Fourteen fractions from 2 HN and 2 rectum patients treated with online adaption were retrospectively analysed. For each adapted fraction the session data was exported as a Dicom-RT series containing the registered CBCT/sCT image pair, external patient contour and deformation vector field (DVF) used to accumulate dose on the planning CT. In-house analyses were completed using a combination of Velocity, Matlab, Plastimatch and 3DSlicer. Qualitatively, each CBCT/sCT pair was displayed as an image blend and scored for registration accuracy. Quantitatively, the agreement between sCT/CBCT was assessed by first converting each pixel value to mass density (MD), masking the images to within the external patient contour and evaluating (i) the 3D Gamma pass rate (criteria ΔMD=3%, DTA=4mm), and (ii) target registration error (TRE). Here TRE was evaluated as the 3D displacement between corresponding intensity-based landmarks detected using the Scale Invariant Feature Transform (SIFT) algorithm. Additionally, the DVF was converted to a Jacobian determinant image and the fraction of negative Jacobian voxels reported. Negative Jacobian voxels were localised within the CBCT/sCT to identify potential problem areas for adaptation. Results For HN patients the visual alignment of anatomic structures between sCT and CBCT appeared to match within 2 voxels (4 mm) and was scored ‘Usable with risk of deformity’. Rectum fractions achieved the same score but with local discrepancies noted in the presence of the bladder and any gas. In terms of the quantitative metrics, MD gamma pass rates showed a tighter range of values for rectum (58- 67%) as compared to HN (53-74%). For the TRE analysis, SIFT identified 50-1800 feature landmarks in each corresponding sCT/CBCT pair. The median (STD) TRE was 2-3mm (3-6mm), with 85% of HN and 90% of rectum landmarks exhibiting TRE<5mm. None of the HN fractions exhibited negative Jacobian values. By comparison each rectum patient had at least 1 fraction associated with a negative Jacobian value. Blended visual inspection localised these non-physical deformations to the bladder and gas in the small bowel. Conclusion We performed an independent evaluation of the sCT used for online plan generation on the Varian adaptive