Transfer learning for automatic sarcopenia evaluation at T12 vertebral level
McSweeney, D Green, Andrew Bromiley, P. A. Van Herk, Marcel Mansoor, Was Weaver, Jamie M McWilliam, Alan
McSweeney, D
Green, Andrew
Bromiley, P. A.
Van Herk, Marcel
Mansoor, Was
Weaver, Jamie M
McWilliam, Alan
Citations
Altmetric:
Abstract
Purpose or Objective
Sarcopenia is a progressive loss of muscle mass and is
emerging as a potential important prognostic factor for RT
patients. It is typically assessed by segmenting skeletal
muscle at the L3 vertebral level and extracting features of
the muscle. However, L3 is not visible on most RT planning
scans.
In this work we use transfer learning to re-train an existing
model, originally for segmenting muscle at L3, to segment
muscle at the T12 vertebral level. The use of transfer
learning significantly reduces the number of training
images required. To show prognostic value, mean skeletal
muscle attenuation (SMA) at L3 and T12 were compared both directly ans in survival models. Material and Methods
An existing model was available, pre-trained to segment
the muscle compartment at L3 (Green et al, ESTRO 2019).
Transfer learning was used to fine-tune the model using
manual delineations at T12 (training n=16, validation n=2,
unseen=10). To ensure no bone remained, a threshold of
226 HU was applied to the CT scan, expanded isotropically
by 0.5 mm and the resulting mask excluded from the
segmentation. Dice score and Distance-To-Agreement
(DTA) were calculated for the unseen patients.
208 oesophago-gastric (OG) cancer patients were available
for testing prognostic value. In this cohort, L3 had been
previously segmented and validated. Segmentations at T12
were generated and visually assessed, to determine the
failure rate. SMA was extracted and compared for patients
with successful L3 and T12 segmentations by calculating
their correlation and performing a paired t-test. Finally,
prognostic value was investigated in Kaplan-Meier curves
(split on median SMA) and multivariate Cox models
including either L3 or T12 SMA controlling for performance
status, age and sex.
Results
Transfer learning with 16 segmentations resulted in a
mean Dice score of 0.75(σ=0.05) and mean DTA of
0.56(σ=0.40) cm in the unseen patients. Of the 208 scans
in the OG cancer cohort 191 segmentations at T12 were
successful (Fig 1). SMA at L3 and T12 were strongly
correlated (R=0.80), but mean SMA was different (25 HU vs
34 HU, p<0.001), Fig 1. Kaplan-Meier curves for L3 and T12
showed significant differences in survival when split on
median density, Fig 2.
Multivariate Cox models identified performance status and
skeletal muscle density as predictive of overall survival.
SMA at both L3 and T12 were found to be prognostic
(p=0.03 and p<0.01, respectively) with hazard ratios 0.98
and 0.97 per HU showing increased muscle density is
beneficial.
Conclusion
We demonstrated a fully automated method for sarcopenia
assessment at T12. Transfer learning with a small training
set resulted in accurate muscle segmentation. Analysis on
a cohort of OG cancer patients shows that SMA at L3 and
T12 are correlated and are both prognostic for patient
outcome. Automating skeletal muscle segmentation at T12
provides a significant step forwards in exploring the
prognostic value of sarcopenia in larger cohorts of patients
treated by RT.
Description
Date
2020
Publisher
Collections
Keywords
Type
Meetings and Proceedings
Citation
Mcsweeney D, Green A, Bromiley PA, Van Herk M, Mansoor W, Weaver J, et al. PD-0541: Transfer learning for automatic sarcopenia evaluation at T12 vertebral level. Radiotherapy and Oncology . 2020 Nov;152:S300.