ESTRO 2021 Abstract Book

S1546

ESTRO 2021

to patient, and from organ to organ for the same patient, suggesting that it reflects an intrinsic, patient related change in the imaging/biology. More elaborate analysis is necessary to better understand the underlying mechanisms of these changes, and to address reproducibility and repeatability of these measurements.

PO-1816 Sarcopenia in lung cancer - population based analysis of skeletal muscle density A. Green 1 , E. Vasquez Osorio 1 , M. van Herk 1 , A. McWilliam 1 1 The University of Manchester, Radiotherapy Related Research, Manchester, United Kingdom

Purpose or Objective Sarcopenia is an age-related degenerative condition where muscle mass and tone is lost. Sarcopenia is prognostic for many cancer sites, however, thresholds are derived from unrelated cohorts or using optimal cut-point methods, known to overestimate significance. Recently, skeletal muscle density (SMD) has been shown to be equally, or more prognostic than skeletal muscle index [1]. In this work, we perform the first large-scale analysis of SMD in a population at risk of lung cancer, evaluating SMD in low-dose CT scans taken as part of the US National Lung Screening Trial (NLST), and produce age- dependant baseline models appropriate for patients at risk of lung cancer. Materials and Methods A standard VGG16-Unet Convolutional Neural Network (CNN) was trained using 85 manual segmentations of the muscle compartment at the level of the 12 th thoracic vertebra. Transfer learning was utilised from the imagenet database, using 68 segmentations for training, 8 for validation and 9 for testing. Dice Similarity coefficient (DSC) measured segmentation performance. After segmentation of the full NLST dataset, muscle characteristics including the segmented area and the SMD were extracted. Area greater than 1,000 image pixels was required to reject failed segmentations. Next, sex-specific summary statistics (median, 5 th centile) were produced, and linear models as a function of age derived. Patients were imaged up to three times in subsequent annual screening visits, each visit was used with the patient’s age incremented by one year subsequent to baseline. Results The CNN achieved a mean DSC in the test set of 0.92 (range 0.88-0.95). Example segmentations from the dataset are shown in Figure 1A. Running on a single Nvidia RTX 2080TI GPU, the network was able to segment all 62,741 slices in 5 hours 40 minutes. Of 62,741 CTs segmented, area thresholding identified 893 failed segmentations (Failure rate of 1.5%). Linear models were created for male sex from 36,489 SMD measurements and for female sex from 25,536 SMD measurements. In the age range of 55-76 y/o, males lose median SMD of 0.57 HUyr -1 , while females lose 0.52 HUyr -1 ; we also observed an offset of 11 HU between the sexes, with the fitted intercept being 63 HU for males, and 52 HU for females. Boxplots showing SMD at each age, overlaid with the linear model relating median SMD to age for each sex are shown in figure 1B.

Conclusion SMD has been analysed in a large population at risk of lung cancer for the first time, showing a strong linear relation with age. Sarcopenia, and in particular SMD, is an important prognostic factor for cancer patients, regardless of site and treatment modality. These models will inform the development of age corrected and risk stratification thresholds of sarcopenia for lung cancer patients. 1: 10.1016/j.clnu.2016.03.010

PO-1817 Modeling response to SBRT/SRS: do indirect damage and saturation play a role? M.A. Gago-Arias 1 , S. Neira 2 , M. Pombar 3 , A. Gómez-Caamaño 4 , J. Pardo-Montero 2

1 Instituto de Invesitación Sanitaria de Santiago, Group of Medical Physics and Biomathematics, Santiago de Compostela, Spain; 2 Instituto de Investigación Sanitaria de Santiago, Group of Medical Physics and Biomathematics, Santiago de