Fast Imaging of the Substantia Nigra with Zero-Shot Super-Resolution
Poster No:
202
Submission Type:
Abstract Submission
Authors:
Vincent Beliveau1,2, Christoph Birkl3, Florian Krismer1, Frank Jagusch1, Ruth Steiger3, Christian Kremser3, Stefan Kiechl1, Elke Gizewski3, Klaus Seppi1, Christoph Scherfler1
Institutions:
1Department of Neurology, Medical University of Innsbruck, Innsbruck, Austria, 2Neurobiology Research Unit, Copenhagen University Hospital Rigshospitalet, Copenhagen, Denmark, 3Department of Radiology, Medical University of Innsbruck, Innsbruck, Austria
First Author:
Vincent Beliveau
Department of Neurology, Medical University of Innsbruck|Neurobiology Research Unit, Copenhagen University Hospital Rigshospitalet
Innsbruck, Austria|Copenhagen, Denmark
Department of Neurology, Medical University of Innsbruck|Neurobiology Research Unit, Copenhagen University Hospital Rigshospitalet
Innsbruck, Austria|Copenhagen, Denmark
Co-Author(s):
Introduction:
The assessment of dorsolateral nigral hyperintensities (DNH) has consistently been shown to provide excellent diagnostic accuracy between patients with Parkinson's disease (PD) and healthy controls (Kim et al., 2019). Although research efforts have been expanded to develop fast and accurate MRI sequences, imaging of the substantia nigra remains challenging and further improvement is required to enable nigral imaging into daily clinical practice. To date, the segmented echo-planar imaging (EPISEG) sequence previously proposed by Hernadi et al. (2021) is the fastest MRI sequence for imaging DNH. In this pilot study, we explored the possibility of reducing the acquisition time of the EPISEG sequence by exploiting redundancy in the acquisition and simultaneously improving imaging quality by using a deep learning super-resolution approach.
Methods:
The EPISEG sequence was acquired for 7 PD patients (64.4 ± 10.8 years; 3 female) and 8 healthy controls (62.4 ± 4.1 years; 1 female) on a 3T MRI scanner. The original sequence consists of 6 measurements acquired over 2:20 min which were averaged online on the scanner to obtain images with an isotropic resolution of 1 mm³. Here, we used the same sequence but each of the 6 measurements were saved independently.
A 3D zero-shot super-resolution (ZSSR) model (Shocher et al., 2018) was trained to double the resolution of the images. The model is a fully convolutional neural network, with 8 hidden layers, each with 64 channels and rectified linear unit (ReLU) activations. Low-resolution images used for training were created by 1) downsampling the original images by a given scale factor using trilinear interpolation and 2) upsampling them back to their original size. Then, the model was taught to predict the residual between the original and low-resolution images. In this way, the model learns to revert the loss of information caused by the downsampling at specific scale factors, a process that can then directly be applied to upsample images. Figure 1 presents an overview of the training and inference. The model was gradually trained to upsample images at increasing scale factors of 1.25, 1.5, 1.75, and 2.
Two approaches were evaluated: 1) the original EPISEG sequence where 6 measurements are upsampled to 0.5 mm³ using trilinear interpolation and averaged, and 2) the proposed approach where 3 out of 6 measurements are upsampled to 0.5 mm³ using the 3D ZSSR model and averaged. The presence of DNH was assessed by a neurologist with 10 years of experience in processing brain MRI. The rater was presented with randomized and anonymized images which were flipped left-right between approaches. Images where both DNH were visible were labeled as "healthy".
A 3D zero-shot super-resolution (ZSSR) model (Shocher et al., 2018) was trained to double the resolution of the images. The model is a fully convolutional neural network, with 8 hidden layers, each with 64 channels and rectified linear unit (ReLU) activations. Low-resolution images used for training were created by 1) downsampling the original images by a given scale factor using trilinear interpolation and 2) upsampling them back to their original size. Then, the model was taught to predict the residual between the original and low-resolution images. In this way, the model learns to revert the loss of information caused by the downsampling at specific scale factors, a process that can then directly be applied to upsample images. Figure 1 presents an overview of the training and inference. The model was gradually trained to upsample images at increasing scale factors of 1.25, 1.5, 1.75, and 2.
Two approaches were evaluated: 1) the original EPISEG sequence where 6 measurements are upsampled to 0.5 mm³ using trilinear interpolation and averaged, and 2) the proposed approach where 3 out of 6 measurements are upsampled to 0.5 mm³ using the 3D ZSSR model and averaged. The presence of DNH was assessed by a neurologist with 10 years of experience in processing brain MRI. The rater was presented with randomized and anonymized images which were flipped left-right between approaches. Images where both DNH were visible were labeled as "healthy".
Results:
By design, the images upsampled with ZSSR exhibited sharper details; see Figure 1 for examples in a healthy control and a PD patient. For the EPISEG approach, the rater identified PD patients and healthy controls with an accuracy of 67% (10/15) (sensitivity=100%; 7/7 and specificity=38%; 3/8), whereas for the images processed with ZSSR, he obtained an accuracy of 80% (12/15) (sensitivity=86%; 6/7 and specificity=75%; 6/8).
Conclusions:
In the original EPISEG sequence, multiple measurements are acquired to average out noise in the images. By acquiring fewer measurements, it is possible to reduce the acquisition time, but, increased noise is correspondingly expected. ZSSR provides an alternative approach to reduce noise and enhance the details of individual measurements. In our evaluation, fewer EPISEG measurements upsampled with ZSSR allowed, in all cases but one, for similar or improved assessment of DNH compared to the original EPISEG sequence. With 3 measurements, the acquisition time for the EPISEG sequence is reduced to a mere 1:10 min, thus making the addition of this sequence to any scanning protocol negligible. More experimentation in larger datasets is required to evaluate this approach thoroughly.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 1
Novel Imaging Acquisition Methods:
Imaging Methods Other 2
Keywords:
Other - Parkinson's disease; Diagnosis; Deep learning; Super-resolution
1|2Indicates the priority used for review
Provide references using author date format
Kim, E. Y., Sung, Y. H., & Lee, J. (2019). Nigrosome 1 imaging: technical considerations and clinical applications. The British Journal of Radiology, 92(1101), 20180842.
Shocher, A., Cohen, N., & Irani, M. (2018). “zero-shot” super-resolution using deep internal learning. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 3118-3126).