Large scale evaluation of sarcopenia as prognostic factor in lung cancer radiotherapy patients

Abstract

Purpose or Objective Sarcopenia is a degenerative condition in which muscle wastes, that has been widely shown to be prognostic for patients treated with chemotherapy. Sarcopenia is also emerging as prognostic factor in radiotherapy. However, to date, cohorts analysed have been small due to the need for manual segmentations. In this work we analysed a very large cohort of non-small cell lung cancer (NSCLC) patients, using an Artifical Intelligence (AI) based automated segmentation to identify skeletal muscle at the third lumbar vertebral level (L3) and demonstrate the prognostic value of skeletal muscle density as a measure of sarcopenia. Material and Methods Whole body PET/CT images from a cohort of 549 NSCLC patients treated with standard fractionation (55 Gy in 20 fractions) were collected. The slices at the center of the L3 vertebral body, manually identified, were segmented using a previously developed AI tool. After visual inspection, the segmentations were used to compute the mean skeletal muscle density (SMD). SMD indirectly measures fat infiltration in the muscle capsule, a common feature of sarcopenia, with lower SMD indicating more fat. Known gender differences in skeletal muscle properties were accounted for by adding +7 HU to SMD for females to equate their median SMD with that of males. An optimal threshold in corrected SMD for survival difference was identified. Kaplan-Meier survival curves were produced for high- and low-SMD groups. A multivariate Cox regression model for overall survival accounting for log tumour size, gender, N stage, Performance Status (PS) and corrected SMD (as a continuous variable) was produced. Results Of the available 549 images, 473 were segmented successfully (figure 1). The majority of failures occurred in poor quality low-dose CT images; likely as the AI tool was trained on higher dose CT imaging. Figure 2a shows the Kaplan-Meier curves produced by splitting the cohort on corrected SMD of 17 HU, the identified optimal threshold for survival difference. A difference in median survival of 3 months is observed where patients with higher SMD do better. In the multivariate Cox analysis SMD remained significant (fig. 2b), with a hazard ratio of 0.99 per HU (p=0.02), indicating that denser muscle is advantageous. In our final model performance status (PS) was not significant. However, without SMD, PS was significant (data not shown). Conclusion We performed semi-automated segmentation for sarcopenia assessment in a very large cohort of lung cancer patients, and it was successful in 86% of the images. We are in the process of improving the AI tool and developing methods to utilize planning thoracic images as alternative. A statistically significant difference in survival was identified for NSCLC patients, where patients with an SMD >17 HU have an additional 3 months median survival. SMD remains significant in multivariate analysis with a hazard ratio of 0.99 per HU. Further work exploring the use of SMD as a quantitative alternative to the qualitative PS is in order.