Jie Ding1, Varut Vardhanabhuti1, Eric Lai2, Yuan Gao3, Sophelia Chan4, and Peng Cao1
1Department of Diagnostic Radiology, Li Ka Shing Faculty of Medicine, The University of Hong Kong, Hong Kong, Hong Kong, 2Li Ka Shing Faculty of Medicine, The University of Hong Kong, Hong Kong, Hong Kong, 3Division of Neurology, Department of Medicine, Queen Mary Hospital, The University of Hong Kong, Hong Kong, Hong Kong, 4Department of Paediatrics and Adolescent Medicine, Li Ka Shing Faculty of Medicine, The University of Hong Kong, Hong Kong, Hong Kong
Synopsis
Time-efficient
thigh muscle segmentation is a major challenge in moving from primarily
qualitative assessment of thigh muscle MRI in clinical practice, to potentially
more accurate and quantitative methods. In this work, we trained a convolutional
neural network to automatically segment four clinically relevant muscle groups using
fat-water MRI. Compared to cumbersome manual annotation which ordinarily takes
at least 5-6 hours, this fully automated method provided sufficiently accurate segmentation
within several seconds for each thigh volume. More importantly, it yielded more reproducible fat fraction estimations, which
is extremely useful for quantifying fat infiltration in ageing and in diseases
like neuromuscular disorders.
Introduction
Fat-water
decomposition MRI provides an excellent approach for non-invasively evaluating
the degree of fat infiltration in muscle tissues. However, the need for
segmentation severely limits its application in quantitative fat assessment in
clinical practice. The conventional methods for segmenting thigh muscles in current
literature[1] rely on manually drawn regions-of-interest
(ROIs), which is an extremely time-consuming and cumbersome process prone to
subjective bias. Some previous studies performed manual segmentation on one or
a few representative slices[2, 3], but such process can potentially
lower the accuracy and reproducibility, which is also not feasible for clinical
adoption. Although recent studies developed semi-automated or automated
algorithms[4, 5], they are still difficult to
apply in clinical settings[6]. A faster, simpler and fully
automated thigh muscle segmentation method for reproducible fat fraction quantification
is highly desired. The purpose of this work is to develop and validate a
convolution neural network (CNN) for automatically segmenting four thigh muscle
groups using a reference database (MyoSegmenTUM)[7]. We further evaluate its
reproducibility in fat fraction estimation by comparing with manual
segmentation. Methods
This
study was based on the reference database MyoSegmenTUM[7], which consists of 25
fat-water decomposition MRI scans collected from 19 subjects using a 6-echo 3D
spoiled gradient echo sequence. The manual segmentation masks of all scans were
provided as ground truth. The manual ROIs delineated four clinically relevant
muscle groups: quadriceps femoris muscle (ROI1), sartorius muscle (ROI2),
gracilis muscle (ROI3) and hamstring muscles (ROI4), with an average
segmentation time of ~6 hours per scan. Three subjects were scanned 3 times
with repositioning, which were used as an independent testing for evaluating
the segmentation accuracy and the reproducibility in fat fraction estimation.
The remaining scans were used to train a CNN for thigh muscle segmentation using
a U-net architecture[8].
For all
scans, the left and right side thighs were first separated for analysis, which
also doubled the dataset size. As a preprocessing step, the subcutaneous fat
was removed automatically on water images by Kmeans clustering followed by an order-statistic
filtering (also remove the skin) and a dilation. The water and fat images covering
only the thigh muscle regions were provided to the CNN as two input channels.
The CNN was trained in 2D using a total of 3808 slices. The training objective
was to minimize the categorical cross‐entropy loss between the CNN outputs and the
manual segmentations.
The
CNN was then tested using the independent set of MRI images acquired from the 3
subjects (6 thigh volumes) with 3 repeated scans. The segmentation accuracy was
evaluated using Dice index. In addition, fat fraction quantification was
performed on the proton density fat fraction (PDFF) maps. Pearson correlation coefficient
was calculated between the mean of PDFF values (meanPDFF) using manual and
automated CNN-derived ROIs of all 1st scans. The reproducibility in
fat fraction estimation was determined by the
intraclass correlation coefficients (ICCs) of meanPDFF using the 3
repeated measures.Results
Figure
1 shows the MRI images and the segmentation results of a representative slice
from the independent testing set. As shown, the manual and automated segmentation
methods have satisfying agreement visually. Table 1 summarizes the mean and
median Dice coefficients between manual and automated segmentations for
ROI1~ROI4. It must be noted that some ROIs are very small, or some regions
(particularly at either ends of muscle insertion/attachment) can be more
inaccurate, which impacts the Dice coefficient. It is thought that Dice coefficient
of 0.7 or more conforms to good agreement for such scenarios[9, 10].
In addition, our median Dice coefficients were >0.8 for all ROIs.
Figure
2 shows that the meanPDFF of all 1st scans calculated using manual and
automated masks were strongly correlated, with Pearson ρ=0.958 (p<<0.0001).
Table 2 demonstrates the ICCs of meanPDFF in the four ROIs. Although both
segmentations yielded moderate reproducibility in ROI2, our CNN-based automated
segmentation method produced very comparable (in ROI3) and more reproducible (in
other ROIs) fat fraction measures compared to manual segmentation.Discussion
In
this work, we adopt the U‐net architecture to train a CNN using fat-water
decomposition MRI for thigh muscle segmentation. Compared to the time-consuming
and labor-intensive manual segmentation, this fully automated segmentation
method is able to delineate four clinically relevant muscle groups with encouraging
accuracy in a few seconds per volume. In terms of fat fraction quantification, the
automated segmentation not only generated fat fraction measures strongly
correlated with those determined by manual segmentation, but more importantly,
demonstrated overall higher reproducibility in fat fraction estimation. Conclusion
As the need for segmentation is a primary challenge
in the quantitative analysis of thigh muscle MRI, the proposed CNN-based
segmentation that provides reproducible fat fraction estimation would be beneficial
for clinical practice such as quantifying fat in muscle associated with ageing,
or in conditions such as neuromuscular disorders. Moreover, a reproducible
measure allows serial monitoring which can be useful particularly if there is
intervention to allow for detection of disease progression or for evaluation of
treatment efficacy. Acknowledgements
No acknowledgement found.References
1. Burakiewicz J, Sinclair CD, Fischer D,
Walter GA, Kan HE, Hollingsworth KGJJon. Quantifying fat replacement of muscle
by quantitative MRI in muscular dystrophy. 2017;264(10):2053-2067.
2. Morrow
JM, Sinclair CD, Fischmann A, et al. MRI biomarker assessment of neuromuscular
disease progression: a prospective observational cohort study.
2016;15(1):65-77.
3. Wokke
B, van den Bergen J, Aartsma-Rus A, Webb A, Verschuuren J, Kan H. Quantitative
assessment of the inter-and intra-muscle fat fraction variability in Duchenne
muscular dystrophy patients. Proceedings of the 19th Annual Meeting of ISMRM;
2011. p. 269.
4. Rodrigues
R, Pinheiro AMJapa. Segmentation of Skeletal Muscle in Thigh Dixon MRI Based on
Texture Analysis. 2019.
5. Mesbah
S, Shalaby AM, Stills S, et al. Novel stochastic framework for automatic
segmentation of human thigh MRI volumes and its applications in spinal cord
injured individuals. 2019;14(5):e0216487.
6. de
Mello R, Ma Y, Ji Y, Du J, Chang EYJAJoR. Quantitative MRI Musculoskeletal
Techniques: An Update. 2019;213(3):524-533.
7. Schlaeger
S, Freitag F, Klupp E, et al. Thigh muscle segmentation of chemical shift
encoding-based water-fat magnetic resonance images: The reference database
MyoSegmenTUM. 2018;13(6):e0198200.
8. Ronneberger
O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image
segmentation. International Conference on Medical image computing and
computer-assisted intervention: Springer; 2015. p. 234-241.
9. Spuhler
KD, Ding J, Liu C, et al. Task‐based assessment
of a convolutional neural network for segmenting breast lesions for radiomic
analysis. Magnetic resonance in medicine 2019;82(2):786-795.
10. Carillo
V, Cozzarini C, Perna L, et al. Contouring variability of the penile bulb on CT
images: quantitative assessment using a generalized concordance index.
2012;84(3):841-846.