Introduction

Multi-slice two-dimensional (2D) turbo spin-echo (TSE) imaging is commonly used because of its high in-plane resolution and relatively short scan time, but its utility is sometimes limited due to poor through-plane resolution ^1,2. Thus, multiple 2D TSE scans are acquired in multiple orthogonal imaging planes for comprehensive image interpretations, increasing the overall scan time ^2–4. In this study, we propose a novel slice-profile-based transformation super-resolution (SPTSR) framework with deep generative networks for super-resolution (SR) of multi-slice 2D TSE imaging. The networks were trained by synthesized low- resolution (LR) training input of coronal scans via slice-profile-based downsampling (SP-downsampling), and the trained networks inferred on the slice-profile-convolved (SP-conv) testing axial scans input for 5.5x through-plane SR.

Methods

Slice-profile Transformations:
Previous studies attempted to synthesize LR reformatted training input via the k-space zero-filling (KS-ZF) downsampling. However, the KS-ZF downsampled LR input from the coronal scan is inherently different from its reformatted version of the axial scan. This is because the coronal scan and the axial scan are convolved with their corresponding slice profiles in their through-plane directions. In this study, we propose to apply slice-profile downsampling (SP-downsampling) to the coronal scans for synthesizing LR training input (Figure 1a), and apply slice-profile convolution (SP-conv) to pre-process axial scans for the networks inference input (Figure 1b).
Deep Generative Networks:
The networks architecture was inspired by SRGAN⁵, with the following three key differences. First, we used three consecutive LR images as input, with the middle slice being the targeted input, to learn the inter-slice spatial relationship. Second, WGAN-GP⁶ training scheme was used, and thus batch normalizations and the last sigmoid activation function were removed from the discriminator network. Third, the upsampling blocks in the generator model were modified to 1D anisotropic upsampling.
MRI Dataset:
We utilized a total of 3,895 clinical subjects with 4,878 pairs of axial and coronal scans using the multi-slice 2D T2-weighted TSE sequence. Both axial and coronal scans were based on the same imaging sequence except for the imaging-plane orientations. Each scan included 20 slices, with an in-plane resolution of 0.625 × 0.625 mm². The slice thickness was 3 mm and the slice spacing was 3.6 mm. The training/validation/testing splits were 3,453/392/50 from 3,895 clinical subjects.
Visual Comparisons:
The proposed SPTSR results from T2w-TSE axial scans in the testing dataset were visually compared with the bilinear interpolation, the baseline KS-ZF trained networks, and SMORE⁷, which is one of the leading self-supervised algorithms for deep learning super-resolution. The corresponding in-plane T2w-TSE coronal scans were served as visual references only, because they were different scans.
Reader Study:
To assess the through-plane SR results, two genitourinary radiologists independently assessed three methods: bilinear interpolation, KS-ZF trained networks (baseline), SPTSR (proposed). In total, 30 test subjects, each with one axial scan, were examined. Three methods were anonymized and randomly shuffled. Along with the visual references (T2w-TSE coronal scans), all four image sets were loaded into OsiriX (Pixmeo SARL, Bernex, Switzerland), and the readers scrolled through the coronal slices. Diagnostic quality metrics of sharpness, artifacts, noise level and overall image quality of each method were scored on a 4-point Likert scale (1 is the worst and 4 is the best). The readers also blindly ranked the overall quality of the three methods. Averaged ratings and rankings from two readers were compared between three methods. Mann-Whitney U tests were used to assess the significant differences (p<0.01) between the three methods. Cohen’s Kappa was calculated for inter-reader variability⁸.

Results

Visual Comparisons:
The examples of the testing results are shown in Figure 2 for visual comparison. Compared to the other methods, the proposed SPTSR output was much less noisy, much sharper, and showed finer detail structure, similar to the visual reference images (T2w-TSE coronal scan).
Reader Study:
The reader study results are as demonstrated in Figure 3. The proposed SPTSR method reaches an overall image quality score of (3.72 ± 0.44), and 20 cases out of 30 received an excellent score (4.0) from both readers. The overall image quality scores from two readers show the decisive advantage of the proposed SPTSR method compared to the other two methods (the proposed method: 29 cases have the best quality; the baseline method: 1 case has the best quality; the BI method: all 30 cases have the worst quality). The ranking results from the two readers are perfectly matched. The proposed SPTSR method is significantly better in all categories compared to the baseline KS-ZF trained network in terms of sharpness, artifact, noise level, and overall image quality (p<0.01). The score results from two readers show substantial agreement with Cohen’s Kappa of 0.68 (95% confidence interval: 0.62-0.74).

Discussions

We proposed a novel slice-profile transformation-based super-resolution (SPTSR) framework for multi-slice 2D TSE MRI. We utilized a large dataset of clinical subjects and scans and demonstrated the visual improvements for 5.5x through-plane SR compared to the k-space zero-filling-based method and the super-resolution algorithm SMORE^7. Our reader study results confirm the superiority of SPTSR compared to the KS-ZF trained networks.

Conclusions

We developed a novel SPTSR framework for super-resolution of multi-slice 2D TSE MRI. The visual results and reader study results demonstrated the effectiveness of our proposed framework in 5.5x through-plane SR.

Acknowledgements

This work was supported by the National Institutes of Health (NIH) R01-CA248506 and funds from the Integrated Diagnostics Program, Department of Radiological Sciences & Pathology, David Geffen School of Medicine at UCLA.