Matt Hemsley1,2, Rachel W Chan2, Liam Lawrence1,2, Sten Myrehaug3,4, Arjun Sahgal2,3,4, and Angus Z Lau1,2
1Medical Biophysics, University of Toronto, Toronto, ON, Canada, 2Department of Physical Sciences, Sunnybrook Research Institute, Toronto, ON, Canada, 3Department of Radiation Oncology, Sunnybrook Health Sciences Centre, Toronto, ON, Canada, 4Department of Radiation Oncology, University of Toronto, Toronto, ON, Canada
Synopsis
qMT has been suggested as a biomarker in Glioblastoma
patients. However, reconstruction involves a computationally expensive fitting
procedure involving the Bloch-McConnell equations. In this work, the use of
neural networks was investigated to perform the fit and to compute uncertainty
heatmaps to identify regions of potential error. The dataset consisted of 164
scans from N=41 glioblastoma patients (33=training, 8=testing). Models were
evaluated using MAE and correlation in the whole-head volumes and specific ROIs.
The model output agreed with a conventional curve-fitting algorithm (r=0.93, and <1% error)
with speed up factors of 240000x. Uncertainty predictions were correlated with prediction
error (r=0.59).
Introduction
Quantitative
magnetization transfer1 (qMT) is a potential biomarker for monitoring
glioblastoma (GBM) response to chemo-radiation therapy, without the need for
contrast administration2,3. A major limitation of qMT is the long fitting
time required to obtain parametric maps. Fast model fitting is a prerequisite
for certain applications, including daily radiotherapy adaptation based on
quantitative MRI. Neural networks have been previously suggested for qMT fitting4,5
but were limited to healthy patients. Our study expands on previous works
by including GBM patients and computes the uncertainty on network predictions. We
also investigated the use of the network outputs as initialization for a
conventional reconstruction procedure, non-linear least-squares fit of the
Bloch-McConnell equations6. Methods
The dataset consisted 41 GBM patients (164 total scans) who
underwent radiation treatment3. Scans were performed on a Philips
Ingenia 1.5T. Four qMT scans were acquired from each patient: before treatment,
at the 10th treatment fraction, 20th treatment fraction,
and at 1-month post-chemoradiation. Patients were grouped for training (n=33)
and testing (n=8). The network input was a set of T1, T2,
B0, B1 maps and 18 MT saturation frequency offsets at 6
different B1 amplitude/duration. Images were downsampled to a
resolution of 5.0x5.0x5.0 mm3. Input features were standardized. Padding
of 1 was added when using the CNN. The outputs were the qMT parameters; the
semisolid fraction (M0b), the exchange rate (R), and the
T2 times for the bound pool (T2b). The output
of the non-linear least-squares fit of the Bloch McConnell equations6
was used as ground truth for network training and evaluation. Model outputs were
compared over the entire brain, in the gross tumor volume (GTV, taken from
clinical treatment planning) and contralateral normal-appearing white matter
(cNAWM, determined with FSL-FAST7). Two network architectures were
investigated: an artificial neural network (ANN), and a convolutional neural network
(CNN). The network architecture and implementation details are described in
figure 1.
Uncertainty predictions were obtained from the networks following
the procedure described by Kendall8. Uncertainty over model
parameters (model uncertainty) was determined using Monte Carlo dropout by
taking the variance of 50 dropout samples. Uncertainty due to noise
in the input data (data uncertainty) was obtained the modifying the mean
squared loss function to include variance, ($$$\sigma^2$$$), $$$MSE_{unc} = (ground truth-output)^2 / \sigma^2 + ln(\sigma^2)$$$ and splitting the final layer of the networks to output a prediction of $$$\sigma^2$$$. We additionally tested using the output of the ANN to
initialize the traditional Bloch-McConnell fit (denoted ‘hybrid’), rather than
using arbitrary values chosen a priori. The runtime of the hybrid method was
compared to the normal fitting runtime (approximately 10 hours). Results
Network training took approximately 30 min for the ANN and 90
mins for the CNN. Once trained, qMT parameters with uncertainty predictions
could be generated in less than one second on both architectures. Using the
output of the ANN on the traditional fit reduced the average run time from
approximately 10 hours per slice to slightly over 3 hours per slice. The mean
absolute error (MAE) of M0b predicted by each model was, ANN=0.001,
CNN=0.006, and hybrid=0.0002 Fig. 2 shows an example output of semisolid
fraction for each method and the difference maps compared to the ground truth.
Fig. 3 shows the correlation and bias of M0b between the ground truth and ANN output
of a slice over the GTV, cNAWM, and the full brain volume using Bland-Altman
analysis. Fig. 4 similarly shows the correlation and bias of ANN uncertainty predictions
the absolute error between predictions and the Bloch fit. Uncertainty was overpredicted,
serving as an upper bound on error.Discussion
qMT fitting on GBM patients performed using an ANN was
accurate compared to a conventional fitting method using the Bloch-McConnell
equations in all three regions of interest, including the GTV. The MAE of M0b between the ANN and Bloch-McConnell
fit was within the standard deviation across the patients for each region3.
In this experiment, the ANN model output was more accurate than the CNN and had
a higher correlation with ground truth over the eight-patient test set. Higher
resolution images may improve CNN performance in future works. For both
networks, the time required to generate qMT images was less than a second, compared
to the 10 hours required to generate the images with least-squares fitting. The
hybrid model was investigated as a more interpretable method that may be more
robust at handling out of distribution input (data the model has not seen
before) compared to the neural networks alone, while still reducing the runtime of
the traditional workflow. This is likely to increase the clinical adoption of
online reconstruction for dose adaptation.
Uncertainty predictions to automatically determine
potentially incorrect voxels for quality assurance is desirable for clinical
implementation. Uncertainty predictions were found to be correlated with errors
between the ground truth and network output. However, uncertainty was overpredicted
compared to the observed error in each ROI. Future work will include
uncertainty calibration to address underpredicted uncertainties.Conclusion
This study demonstrated that qMT fitting using neural
networks was accurate when compared to traditional non-linear least-squares
fitting. The increased efficiency and associated uncertainty predictions are
expected to facilitate clinical implementation of qMT as a biomarker for GBM
progression. Acknowledgements
We gratefully acknowledge funding from NSERC (RGPIN-2017-06596) and NVIDIA Corporation with the donation of the GPUs used in this work.References
1.
Kucharczyk W, Macdonald P.M,
Stanisz G.J, Henkelman R.M. Relaxivity and magnetization transfer of white
matter lipids at MR imaging: Importance of cerebrosides and pH. Radiology.
1994;192:521–529.
2.
Mehrabian H, Myrehaug S, Soliman H, Sahgal A, Stanisz GJ.
Quantitative Magnetization Transfer in Monitoring Glioblastoma (GBM) Response
to Therapy. Sci Rep. 2018;8(1):2475.
3.
Chan R. W, Chen H, Myrehaug S, Atenafu E. G,
Stanisz G. J, Stewart J, Maralani P. J, Chan A. K. M, Daghighi S, Ruschin M,
Das S, Perry J, Czarnota G. J, Sahgal A, Lau A. Z. Quantitative CEST and MT at
1.5T for monitoring treatment response in glioblastoma: early and late tumor
progression during chemoradiation. Journal of Neuro-Oncology. 2020;1–12.
4.
Luu H. M, Kim D, Kim J, Choi S, Park S. qMTNet:
Accelerated quantitative magnetization transfer imaging with artificial neural
networks. Magnetic Resonance in Medicine. 2021;85(1): 298–308.
5. Kim
B, Schär M, Park H. W, Heo H. Y. A deep learning approach for magnetization
transfer contrast MR fingerprinting and chemical exchange saturation transfer
imaging. NeuroImage. 2020; 221:117165.
6. Chan
R. W, Myrehaug S, Stanisz G. J, Sahgal A, Lau A. Z. Quantification of pulsed
saturation transfer at 1.5T and 3T. Magnetic Resonance in Medicine. 2019;82(5);1684–1699.
7.
Zhang Y, Brady M,
Smith S. Segmentation of brain MR images through a hidden Markov random field
model and the expectation-maximization algorithm. IEEE Trans Med Imag. 2001;20(1):45-57.
8.
Kendall A, Gal Y. What uncertainties do we need in Bayesian
deep learning for computer vision? Adv. Neural Inf. Process. Syst. 30.
2017;5580–5590
9.
Paszke A. et al. PyTorch: An Imperative Style,
High-Performance Deep Learning Library. Adv. Neural Inf. Process Syst.
32. 2019;8024–8035.