Sibaji Gaj1, Ceylan Colak2, Mingrui Yang1, Kunio Nakamura1, and Xiaojuan Li1
1Department of Biomedical Engineering, Cleveland Clinic, Cleveland, OH, United States, 2Department of Radiology, Cleveland Clinic, Cleveland, OH, United States
Synopsis
Synovitis
is a very common finding in joints of RA patients, which may serve as
biomarkers for early diagnosis and for early treatment response evaluation. However,
synovitis quantification is challenging because manual segmentation of such
irregular lesions is tedious and prone to inter reader variation. In this work,
we implemented a fully automatic segmentation algorithm for synovitis lesions
in wrist Magnetic Resonance images in subjects with RA using deep learning
based conditional generative adversarial networks and U-Net. Using a small
number of training data, the proposed model demonstrated feasibility of fully
automatically synovitis segmentation with reasonable accuracy (Dice coefficient
0.78).
Introduction
Rheumatoid
arthritis (RA) is a chronic inflammatory disease affects approximately 1% of
the US population, caused by destruction of articular and periarticular
structures [1]. The early diagnosis and accurate monitoring of rheumatoid
arthritis (RA) are essential to delay joint destruction and functional
disability and also necessary for evaluating successful therapy. Magnetic
Resonance images (MRI) is a sensitive tool for the detection and quantification
of synovitis [2]. However, the transition of quantitative MRI in clinical settings
requires fast accurate fully automatic segmentation. In this work, we have
presented a fully automatic segmentation algorithm for synovitis in wrist MR in
subjects with RA using deep learning-based conditional generative adversarial
networks (cGAN) [3] and U-Net [4].Method
26 patients with RA were studied at baseline (n=17),
1-month (n=6), 3-month (n=19) and 1-year (n=10), resulting a total of 52 exams.
MRI data were collected on a 3-T scanner (GE Healthcare) with a 16Rx wrist coil
(InVivo). The imaging protocol included pre- and post-Gd T1-weighted spoiled
gradient echo (SPRG) images (dimension 512x512x20, in plane resolution 0.4mm,
2mm slice thickness. The data were split randomly into 42 (21 subjects) : 10 (5
subjects) for training: testing set. Deep learning architecture based on CGAN was
used where two networks are trained simultaneously: one focuses on generating
realistic segmentation (generator) and the other discriminating between the
manual segmentation and the generated one (discriminator). 2D-UNet with 10 convolution layers of encoder
and decoder was used as generator and another typical convolutional neural
network with 10 layers was used as a discriminator. The generator UNet took 2D slices with two
channels (T1-pre and post contrast) as input and provides pixel wise
probability map for synovitis. The output probability map along with two
channels was provided to the discriminator for segmentation feedback. Along
with adversarial loss from discriminator, Dice coefficient loss between manual
and auto segmentation, and the feature loss from the discriminator were used to
train the network. Adam optimizer was used with an initial learning rate of
10e-4. Batch size was 10. In training,
the input MRI volumes were augmented by random flip along X-axis and Y-axis
during runtime. The model was implemented in python using Keras 2.2.2 (5) and
Tensorflow 1.10.0 [6] framework and trained on Owens Cluster with NVIDIA Tesla
P100 GPU of Ohio Supercomputer Center [7]. The segmentation performance was
evaluated using the Dice coefficient comparing the automatic segmentation and
the manual segmentation. Results
An average Dice coefficient of 0.78 was obtained with
standard deviation of 0.08 across a test set of 10 volumes, compared with
manual segmentations. The example of the automated segmentations along with
manual segmentation is shown in Fig 1. The correlation between volumes and
intensity of manual segmentation and automatic segmentation are R=0.77 and R=0.87.Discussion
Synovitis is a very common finding in joints of
RA patients, which may serve as biomarkers for early diagnosis and for early
treatment response evaluation. However, synovitis quantification is challenging
because manual segmentation of such irregular lesions are very time consuming
and fully automated methods are desirable. Using a small number of training
data, the proposed model demonstrated feasibility of fully automatically synovitis
segmentation with reasonable accuracy. In Fig 1A, it can be observed that the
auto segmentation has learned the general segmentation regions and can segment the
cases even if the structure is irregular due the bone erosion (Fig 1B). Though,
model sometimes over segment bright regions within the bone as synovitis (pointed
by arrow depicted in Fig 1C) and fails to segment lesions outside the regions (pointed
by arrow depicted in Fig 1D). Conclusion
In this study, we presented a deep
learning-based approach to automatically segment synovitis lesions in a patient
with RA using CGAN-UNet based network. The proposed network can obtain
reasonable segmentation performance using very few manual segmentations and can
segment small synovitis lesions. In future work, we will improve performance by
including more manual segmentation, other image contrasts such as T2-weighted
image and optimizing model parameters. Also, the simultaneous segmentation of
other types of lesions such as bone marrow edema will be incorporated in the
model.Acknowledgements
The data collection was supported by UCB Pharma. References
[1] Mikuls T., Saag K. Comorbidity in rheumatoid
arthritis, Rheumatic Disease Clinics of North America, Volume 27, Issue 2,
2001(283-303)
[2] Yang H1, Rivoire J, Hoppe M, Srikhum W, Imboden
J, Link TM, Li X. Computer-aided and manual quantifications of MRI
synovitis, bone marrow edema-like lesions, erosion and cartilage loss in
rheumatoid arthritis of the wrist. Skeletal Radiol. 2015 Apr;44(4):539-47.
[3] Liu F. SUSAN: segment unannotated image
structure using adversarial network. Magn Reson Med. 2019 May;81(5):3330-3345.
[4] Norman B, Pedoia V, Majumdar S. Use of 2D U-Net
Convolutional Neural Networks for Automated Cartilage and Meniscus Segmentation
of Knee MR Imaging Data to Determine Relaxometry and Morphometry. Radiology.
2018;288(1):177-185.
[5] Chollet, F. and others (2015) keras,
https://keras.io.
[6] Abadi M, Barham P, Chen J, et al. TensorFlow: a
system for large‐scale machine learning. In Proceedings of the 12th USENIX
conference on Operating Systems Design and Implementation, Savannah, GA, 2016.
p. 265‐283.
[7] Ohio Supercomputer Center.
http://osc.edu/ark:/19495/f5s1ph73.