Content uploaded by Soumick Chatterjee
Author content
All content in this area was uploaded by Soumick Chatterjee on Sep 28, 2021
Content may be subject to copyright.
118
DDoS: Dynamic Dual-Channel U-Net for Improving Deep Learning Based Super-Resolution of Abdominal
Dynamic MRI
SarasaenC.1,2, ChatterjeeS.1,3,4, NürnbergerA.3,4,5, SpeckO.1,5,6,7
1Otto von Guericke University Magdeburg, Biomedical Magnetic Resonance, Magdeburg, Germany 2Otto von Guericke University Magdeburg, Institute for Medical Engineering, Magdeburg, Germany 3Otto von Guericke University Magdeburg, Faculty of Computer Science, Magdeburg, Germany 4Otto von Guericke University Magdeburg, Data and Knowledge Engineering Group, Magdeburg, Germany 5Center for Behavioral Brain Sciences, Magdeburg, Germany 6German Center for Neurodegenerative Disease, Magdeburg, Germany 7Leibniz Institute for Neurobiology, Magdeburg, Germany
Introduction:
The trade-off between spatial and temporal resolution in dynamic MRI is a hindrance for MR-guided
interventions, which require high temporal resolution while visualizing details. Deep learning based super-
resolution (SR) has shown promising results in dealing with this trade-off[1]. Nevertheless, the available
temporal information of dynamic MRI has not been exploited in this prior work. The potential of improving
the reconstruction quality of dynamic MRIs by incorporating the temporal information has been
demonstrated recently[2-4]. This work extends the previous work by ameliorating the previous model[1] a
the dual-channel (static+dynamic images) super-resolution approach, termed DDoS (Dynamic Dual-channel
of SuperRes).
Methods:
An artificial dynamic dataset was generated by applying random elastic deformations[5] to the publicly
available CHAOS dataset (T1 in- and opposed phase)[6]. Then, a low resolution dataset was simulated by
performing in-plane undersampling[7,8] taking only the centre of the k-space of each slice. A modified U-
Net model[1] was trained with dual-channel input, consisting of the low-resolution image of the current time
point (LR_TPn) and the high resolution image of the previous time point (HR_Tpn-1), see Fig.1. This
training strategy tries to let the network learn the spatio-temporal relationship over time points, with the help
of perceptual loss[9] using a perceptual loss network[10], and was minimised using Adam optimiser for 18
epochs. 3D patches of the volumes were created for training with a patch size of 243, and with a stride of 6
for the slice dimension and 12 for the rest. For testing, a stride of 3 was used for all dimensions with the
same patch size. A 3D abdominal dynamic MRI data was acquired at 3T Siemens Magnetom Skyra [GRE,
T1w Flash 3D, TR: 2.23ms, TE: 10.93ms, voxel size: 1.09x1.09x4.0mm] and was also undersampled in the
same manner and were used for testing.
Results and Discussion:
The highest undersampling (only 6.25% of centre k-space per slice) explored in the former work[1], was
applied to evaluate the improvements with the proposed approach. Fig.2 shows the comparison of low-
resolution input (for the lowest resolution examined), SR result after subject specific fine-tuning[1], DDoS
results and its corresponding ground-truth images. The average SSIM value while comparing the results
against the ground-truth improved from 0.949±0.003 (for SR result after fine-tuning[1]) to 0.979±0.004 and
the average PSNR value improved from 34.647±0.240 to 39.411±0.536. One can observe qualitatively that
the results of the DDoS model are more similar to the ground-truth compared to the result of SR after fine-
tuning. In conclusion, this research illustrates that by incorporating the temporal information in a super-
resolution approach[2] using the DDoS model can improve the reconstruction quality, and may be extended
for application during real-time interventions due to the fast inference speed.
Fig. 1
Fig. 2
Fig. 3