Background

Glioblastoma multiforme (GBM) is an aggressive brain malignancy that is difficult to prognosticate. Sarcopenia is associated with poor survival in cancer and is assessable on cross-sectional imaging, but manual muscle quantification is time-consuming and susceptible to interrater inconsistency. The aim of this study was to develop a fully-automated deep learning-based system for temporalis muscle quantification, a sarcopenia surrogate on cranial MRI, and determine whether it predicts survival in GBM.

Method

Retrospective cohort study of 152 MR studies in 45 patients with newly-diagnosed GBM from 01/2015 - 05/2018. A neural network for automated temporalis segmentation was trained and validated on three loss functions, and used to quantify muscle cross-sectional area (CSA). The primary outcome was effect of temporalis CSA on overall (OS) and progression-free survival (PFS) in GBM. Secondary outcomes were OS and PFS by CSA in age-/sex-disaggregated subgroups and at different longitudinal timepoints.

Results

Mean (SD) age was 54.3 ± 11.0 and 34/45 (75.6%) patients were males. The model achieved high accuracy in segmenting temporalis with Dice coefficient of 0.912 ± 0.031, Hausdorff distance of 1.813 ± 0.390mm and CSA error of 1.94 ± 8.86%. Dice loss conferred best model performance. CSA was a significant predictor of OS (HR 0.506, 95%CI 0.267-0.959; p=0.037) and PFS (HR 0.348, 95%CI 0.182-0.667; p=0.001). CSA is an independent prognostic factor; in multivariate analysis, HR was 0.456 for OS (95%CI 0.207-1.004; p=0.051) and 0.309 for PFS (95%CI 0.131-0.730; p=0.007). The effect of CSA on OS/PFS is independently significant and particularly strong in younger patients (HR 0.213, 95%CI 0.058-0.788; p=0.020/HR 0.053, 95%CI 0.009-0.302; p=0.001) and males (HR 0.400, 95%CI 0.196-0.986; p=0.046/HR 0.173, 95%CI 0.059-0.509; p=0.001), and persists at the post-treatment timepoint.

Conclusion

Temporalis muscle area is a significant prognostic marker in GBM, and can be automatically, rapidly and accurately assessed using a deep learning-based system.

Impact statement

This is the first study, in any cancer type, to apply deep learning to muscle segmentation for sarcopenia assessment and demonstrate prognostic significance, establishing a role for deep learning in sarcopenia screening in cancer, which adds significant value to routine imaging and has clinical application for risk profiling, treatment stratification and interventions for muscle preservation.