Francesco Caliva1, Adam Noworolski2, Andrew Leynes1,3, Claudia Iriondo1,3, Sharmila Majumdar1, Peder Larson1, and Valentina Pedoia1
1Department of Radiology and Biomedical Imaging, University of California, San Francisco, San Francisco, CA, United States, 2EECS, University of California, Berkeley, Berkeley, CA, United States, 3Department of Bioengineering, University of California, Berkeley, Berkeley, CA, United States
Synopsis
We
propose a novel task based deep learning framework for simultaneous MRI reconstruction
and segmentation. On a dataset of retrospectively undersampled knee-DESS
volumes we demonstrate that irrespective of ultra-high acceleration factors
(i.e. 48×) a multitask 3D encoder-decoder is capable of reconstructing with
high fidelity the knee MRI, accurately segment cartilaginous and meniscal
tissues and reliably provide cartilage thickness. Our multitask solution outperforms
two other methods: a compressed sensing reconstruction step, followed by a deep
learning-based tissue segmentation. The other method comprises a cascade of two
convolutional neural networks that sequentially perform image reconstruction
and segmentation.
Introduction
Long MRI scan times lead to low patient throughput, problems with
patient comfort, artifacts from patient motion, and high exam costs1. Many methods have
been proposed to accelerate MRI. However, there is a need for greater systematic
evaluation of the impact of accelerated-MRI on the extraction of clinically
relevant metrics. Furthermore, fast image acquisition and accurate image
post-processing are typically considered as two independent problems. In this
study, we propose a multi-task deep learning approach capable of simultaneously
reconstructing accelerated MRI (up to 48×AF) and segmenting tissues of
interest. As example, we apply our method on retrospectively accelerated-knee
MRI, where we segment tibial, femoral and patellar cartilage and menisci.
Finally, we show that irrespective of the AF, our method can reliably measure
knee cartilage thickness.Methods
172 3D Dual-Echo-Steady-State (DESS) knee volumes (TE 4.7 ms,
TR 16.2 ms, field of view 14 cm,
matrix 307 × 384, slice thickness 0.7 mm, and
bandwidth 185 kHz) were used to evaluate the acceleration methods. Volumes
were Fourier Transformed to generate synthetic k-space data and retrospectively
undersampled by using a variable-density Poisson disk undersampling pattern2 with AF=[1.5, 2, 3, 4, 6, 12, 24, 36, 48] along the encoding phase directions. To
simulate a compressed sensing acquisition and reconstruction (L1-CS)3,
all the images were reconstructed via constrained optimization with an
L1-Wavelet regularization using a primal-dual hybrid gradient algorithm4,5. After image reconstruction, cartilage and meniscus were automatically
segmented using a 3D V-Net6
on all the L1-CS knee volumes. We refer to the cascade of L1-CS reconstruction
and V-Net segmentation as “CS-DL”. The same undersampled dataset was also reconstructed
and subsequently segmented using two identical and independent V-Nets: one trained to denoise images
obtained from zero-filled Inverse-Fast Fourier Transform of the undersampled data and one trained to
segment the output of the first network. We refer to this cascade of V-Nets
with the term “cascadeRS”.
To devise a task-based MRI
reconstruction framework we propose a multitask deep learning “MTDL” approach. Fig.1 is descriptive of a novel deep
neural network architecture we propose. Overall the network consumes a
zero-filled k-space under-sampled MRI volume, and outputs a reconstructed MRI
in addition to the tibial, femoral, patellar cartilage segmentation maps. The
network has the structure of a 3D fully convolutional encoder-decoder, with an encoding
path shared between tasks. Once a common embedded representation of both tasks
is encoded, the network branches into two bottom level paths where task
specific feature representations are learnt. As additional novel element, the
network presents skip connections between the encoding and the image
reconstruction paths, as well between the reconstruction and the segmentation
paths. The reasoning behind such architectural design decision lies in the
quality of the features that are passed through the network’s encoding-decoding
paths via skip-connections: the features available at the encoding path suffer
from under-sampled k-space artefacts, which, especially at high acceleration
factors, result in a severe loss of finer details, which are crucial for tissue
segmentation. On the contrary, the features available in the down sampling path
are instrumental for reconstructing a high-resolution MRI, they provide a good
initial solution, resulting in a faster convergence compared to a gaussian
noise like initialization. The flow of features between the decoding paths,
provides the segmentation branch with features representative of finer details.Results
We tested our
methods on a test-set of 28 volumes. Fig.2 is exemplary of the
performance of MTDL on the reconstruction task at 6× and 12× AF. Fig.3 shows
the reconstruction and corresponding cartilage segmentation and cartilage
thickness maps obtained using the compressed sensing algorithm (A), the
reconstruction portion of cascadeRS (B), and the output of the
reconstruction branch of MTDL (C) on a 48× AF MRI. Fig.4-bottom
depicts averages and 95C-I Dice Score Coefficient (DSC) as function of all
the AF for the three methods. Up to AF 6×, all the three reconstruction
techniques provide good enough quality to achieve a good tissue segmentation
with DSC of: femur 0.81±0.04, 0.84±0.03, 0.87±0.02; tibia 0.81±0.06, 0.84±0.04,
0.86±0.04; patella 0.76±0.09, 0.79±0.07, 0.81±0.08; menisci
0.79±0.05, 0.82±0.03, 0.83±0.03. Conversely, while cascadeRS and MTDL
maintain high quality segmentation up to 48×; from AF 12×, CS-DL reveal a
noticeably DSC drop: femur 0.69±0.05, 0.84±0.03, 0.87±0.02 tibia
0.70±0.11, 0.83±0.08, 0.86±0.03; patella 0.60±0.15, 0.79±0.08, 0.82±0.08;
menisci 0.70±0.10, 0.82±0.03 and 0.84±0.03 respectively. Violin plots in
Fig.4-top report the outcome of a t-test conducted to analyse
whether MTDL significantly outperforms cascadeRS at ultra-high AF. Fig.5 shows
a Bland-Altman plot where the cartilage thickness measured on the manually
annotated segmentation, is compared against the thickness computed on the
automatically MTDL segmentation, showing that there is high correlation, no
bias and no significant difference between the two measurements. Discussion and Conclusions
In this retrospective study, we show that the data-driven nature of DL based
solutions have the potential to make task-based ultraFast MRI feasible. Our
multi-task optimization framework proved to be a convincing solution, it uses
fewer trainable parameters than conventional cascades of Deep Learning
approaches. This was reinforced by a statistically significant improved
performance over a commonly used compressed sensing reconstruction technique.
Our future work will involve testing the framework on real raw data, as well as
further investigate the potential application of ultrafast MRI. Acknowledgements
This project was supported by R00AR070902 (VP),
R61AR073552 (SM/VP) from the National Institute of Arthritis and
Musculoskeletal and Skin Diseases, National Institutes of Health, (NIH-NIAMS).References
1 Zbontar, J., et al.
"fastmri: An open dataset and benchmarks for accelerated mri." arXiv
preprint arXiv:1811.08839 (2018).
2 Lustig, M., et al, “Sparse MRI: The application of compressed
sensing for rapid MR imaging,” Magn. Reson. Med., vol. 58, no. 6, pp.
1182–1195, Dec. 2007.
3 Hammernik, K., et al.
"Learning a variational network for reconstruction of accelerated MRI
data." Magnetic resonance in medicine 79.6 (2018): 3055-3071.
4 Bridson, R.. “Fast
Poisson disk sampling in arbitrary dimensions.” SIGGRAPH sketches. 2007.
5 Chambolle, A., & Pock, T. (2011). A first-order
primal-dual algorithm for convex problems with applications to imaging. Journal
of mathematical imaging and vision, 40(1), 120-145.
6 Tamir J, et al.,
“Parameter-free Parallel Imaging and Compressed Sensing”, Proc. Intl. Soc. Mag.
Reson. Med. 26 (2018).
7.
Milletari, F., et al.
"V-net: Fully convolutional neural networks for volumetric medical image
segmentation." 2016 Fourth International Conference on 3D Vision (3DV).
IEEE, 2016.