Introduction

Fat-suppressed (Fs) contrast-enhanced (CE) three-dimensional (3D) T1-weighted imaging (T1WI) enables clear visualization of head and neck structures. However, this sequence sometimes needs long scanning times which can cause the patient discomfort and even pain. Recently, deep learning (DL)-based image reconstruction techniques, particularly using convolutional neural networks (CNNs), have shown promising results for accelerating the data acquisition process of MRI. Particularly, the advanced DL reconstruction technique of model-based reconstruction using Adaptive-CS-Net, named SmartSpeed AI, have been introduced effective for denoising by combining the CNN and the image reconstruction process of the compressed-sensing sensitivity-encoding (CS)-based denoising cycle¹. The aim of this study was to assess the utility of SmartSpeed AI, by comparing among three reconstruction technique of 1) conventional CS, 2) CS followed by end-to-end DL reconstruction, and 3) model-based DL reconstruction (i.e., SmartSpeed AI) in Fs-CE-3D T1WI of the head and neck.

Methods

The protocol of this retrospective study was approved by our institutional review board.
Patients: Twenty-four patients with head and neck tumor were included in this study. All patients underwent MR scanning by a 3.0-Tesla MR unit (Ingenia Elition X; Philips Healthcare, Best, Netherlands) with a 16-channel neurovascular coil.
Image acquisition: The Fs-CE-3D T1WI of the head and neck were acquired as follows: a 3D-T1 fast-field echo (T1-FFE) sequence, TR 5.9 ms, TE 1.2 ms, flip angle 8 degree, acquired matrix 300 × 300 (reconstructed matrix, 640 × 640), field of view (FOV) 240 × 240 mm, pixel size 0.375 × 0.375 mm, slice thickness 0.8 mm, slice pitch 0.4 mm (total 500 slices), reduction factor 12, scanning time 1 min 36 s. By using the same raw-image dataset, three types of images were reconstructed: 1) conventional CS reconstruction, 2) conventional CS followed by end-to-end DL reconstruction, and 3) SmartSpeed AI. End-to-end DL reconstruction mainly consists of the denoising convolutional neural network (DnCNN) for image processing². SmartSpeed AI utilized a reconstruction model incorporating the DL architecture of Adaptive-CS-Net into the CS-based iterative denoising cycle, described as model-based type DL reconstruction^1,3,4.
Image analysis: Qualitative image assessment was performed by the board-certified radiologists who specialized the head and neck imaging in blinded fashion. The overall image quality, visualizations of anatomical structures, and degree of artifacts were visually evaluated based-on the following five-point (1-5; 1 very poor, 2 poor, 3 moderate, 4 good, 5 excellent) grading system. Quantitative assessment was also performed by placing the region of interests (ROIs) on the tumor and posterior neck muscle to measure the value of signal to noise ratio (SNR). The SNRs were calculated as the mean signal in the ROI divided by the standard deviation (SD) of the signal in the ROI.
Statistical analysis: Statistical comparisons were performed using the Wilcoxon signed-rank test for qualitative analysis and the paired t-test for quantitative analysis. A p value <0.05 was considered significant, with Bonferroni correction of p-value for multiple comparisons among three types of reconstruction techniques.

Results

Twenty-four patients had primary head and neck tumors as follows: pharyngeal squamous cell carcinoma (SCC) (n=10), oral cavity SCC (n=4), nasal or sinonasal cavity SCC (n=4), and parotid gland tumor (n=6).
In the qualitative analysis, scores in SmartSpeed AI-based images were significantly higher than other two types of reconstruction techniques in all evaluation items (p<0.05, respectively). In addition, scores in CS followed by end-to-end DL reconstruction-based images were significantly higher than CS reconstruction in all evaluation items (p<0.05, respectively), except the lesion edge sharpness. Details of qualitative analysis are presented in Fig 1.
In the quantitative analysis, all SNRs in SmartSpeed AI-based images were significantly higher than other two types of reconstruction techniques (p<0.05, respectively). In addition, all SNRs in CS followed by end-to-end DL reconstruction-based images were significantly higher than CS reconstruction (p<0.05, respectively). Details of qualitative analysis are presented in Fig 2.
The representative case is presented in Fig 3.

Discussion

Our results indicated that model-based DL reconstruction (i.e., SmartSpeed AI) consists of the CNN architecture embedded into the CS denoising cycle successfully demonstrated improved image quality in both qualitative and quantitative assessments of the Fs-CE-3D T1WI for evaluation of head and neck compared to the conventional CS and CS followed by end-to-end DL reconstruction technique. We speculated that the algorithm in SmartSpeed AI successfully demonstrated image processing with large amounts of processed data in iterative process through the Adaptive-CS-Net for more accurate denoising^3,4.

Conclusion

Smartspeed AI has the potential to achieve clinical usefulness by providing high image quality within a short scanning time for the detailed evaluation of the head and neck.

Acknowledgements

None