ESTRO 2021 Abstract Book

S966

ESTRO 2021

It has been proved that the incidence of cardiovascular events and survival rates in patients with locally advanced non-small cell lung cancer that are treated with chemotherapy and radiotherapy with radical intention is linked to the cardiac dose. Patients with cardiovascular risk factors may have a higher incidence of adverse events related to their comorbidity. Studying the relationship between the dose of radiotherapy received by the heart and the cardiovascular risk factors with the incidence of cardiac events and survival rates. Materials and Methods Retrospective study in 180 patients diagnosed with locally advanced NSCLC treated with 3D-radiotherapy and chemotherapy with radical intention between 2009 and 2017. Variables to study: 1. Radiotherapy: Pulmonary mean dose, Pulmonary V20, Cardiac V5, Cardiac V30. 2. Cardiovascular Risk Factors: SAT, Total cholesterol levels, Cardiovascular SCORE (Sans et al 2007). Statistic analysis: 3. Actuarial Survival using the Kaplan Meier Method. 4. Univariate Analysis with Log-Rank Test and Multivariate Analysis with Cox Regression of the relationship between the variables under study and survival rates. Results Outstanding results of the analysis have been: • Mean follow-up: 32.5 months. • Mean overall survival: 17.1 months. • Mean disease free survival: 12.8 months. • Univariate analysis for cardiovascular events: Age with statistically significant(p=0.001). Charlson index with statistically significant(p=0.022). Blood pressure with non-statistically significant(p=0.06). Pulmonary V20 with statistically significant(p=0.026). Cardiac V5 with statistically significant(p=0.05). • Multivariate analysis for cardiovascular events: Age with a hazzard ratio of 1.08 (1.004-1.148) (p=0.002). Pulmonary V20 with a hazzard ratio of 1.08(1.037-1.127) (p=0.0001). Conclusion By determining the relationship between the dose of radiotherapy received by the heart, cardiovascular risk factors and cardiac events, we could identify high risk patients for whom the dose received by the organs at risk should be minimised but without compromising PTV coverage or the oncological outcome. PO-1164 Clinical evaluation of an interactive deep-learning assisted contouring method for target contouring M.J. Trimpl 1,2 , P. Charlton 3 , S. Teoh 3,4 , H. Zeng 5 , K.A. Vallis 3 , E.J. Stride 1 , M.J. Gooding 2 1 University of Oxford, Institute of Biomedical Engineering, Oxford, United Kingdom; 2 Mirada Medical Ltd, Science, Oxford, United Kingdom; 3 University of Oxford, Oxford Institute for Radiation Oncology, Oxford, United Kingdom; 4 Oxford University Hospitals, NHS Foundation Trust, Oxford, United Kingdom; 5 Maastrict University Medical Centre+, Department of Radiation Oncology, Maastricht, The Netherlands Purpose or Objective To evaluate the impact on contouring time for target contouring when using an interactive deep-learning assisted contouring tool as compared to standard manual contouring. Materials and Methods The time spent contouring the GTV was compared for 2 contouring tools applied to CT imaging datasets of 10 non-small-cell lung tumours, each contoured by 3 clinicians. The manual and deep learning tools shared the same user interface and a typical basic tool set for contouring. With standard manual segmentation, linear interpolation between two contours on axial image slices enabled estimation of contours on intermediate slices. In contrast, the deep learning tool could predict contours of the structure to be segmented on adjacent slices to an image slice with an existing contour. The resulting contours could subsequently be edited by the clinicians. All user interaction was automatically tracked. The drawing and editing of contours by dragging the cursor was recorded as active contouring time. All other behaviour was logged as observation time. After contouring a case using one contouring method, the same case was contoured one week later using the other method. The order in which manual and deep learning tools were used was randomized per case and per clinician to mitigate the impact of order on the overall timing statistics. Results Using the interactive deep learning tool led to a mean decrease in active contouring time of 33% relative to the standard manual segmentation (Fig. 1). Observation time was decreased by 26%. The reduction in contouring time was significant for both active contouring time and observation time (p<0.001). The change in active contouring time directly reflects the impact of the deep learning tool. The change in observation time can be attributed to multiple factors, such as the time it takes to identify the tumour, to change between contouring tools and to do a final review of a scan. Observation time made up nearly 80% of interaction time for both contouring approaches. Assessment of the workflow timing (Fig. 2) for both segmentation methods showed that although the predictions of the deep learning tool required editing in 68% of the slices, the edits to the predicted contours required substantially less time as compared to the standard manual method. The total contouring time varied greatly for each tumour volume. On average the time spent contouring per case was reduced from 13min to 9min when using the deep learning tool.