Synopsis

Keywords: Skeletal, Bone

The integration of compressed-SENSE and artificial intelligence allows to acquire high-resolution images within relatively short scan times The purpose of this study is to compare the image quality of the musculoskeletal images reconstructed with Compressed -SENSE (CS), CS with integrated artificial intelligence (CS-AI) and CS-AI combined with superresolution(CS-AI-HR) . We found that the image quality was significantly improved by using CS-AI-HR compared to CS. This study provides supportive information for the application of CS-AI-HR in routine clinical practice.

Introduction

Equipment efficiency is a constant challenge for imaging devices, which mainly balances two aspects: speed and image quality. With traditional acceleration techniques, speed is increased, but at too high acceleration factors image quality can be severely jeopardized. High-resolution MRI images can provide excellent anatomical information for precise diagnosis and for treatment guidance. However, high-resolution image scanning requires considerable time, which may cause patient discomfort and, potentially, motion.
In order to reduce the scan time without compromising the image quality, a series of techniques represented by Compressed-SENSE (CS) have been implemented in the imaging field. Recently, the technique of applying artificial intelligence (AI) to CS for acceleration, named as CS-AI, has been introduced¹. CS-AI can increase the acceleration multiplier more than conventional CS while obtaining the same image quality. Its advantage has been demonstrated in the imaging of the ankle² and brain^3,4.
Also, superresolution AI models have been introduced to improve the sharpness of images. The combination of superresolution with acceleration based on CS-AI promises to acquire high resolution images from highly undersampled k-sapce data. This combination is denoted here as CS-AI-HR. The purpose of this study is to compare the image quality of the musculoskeletal images reconstructed with CS, CS-AI and CS-AI-HR.

Methods

5 subjects were included in this study. Routine MR sequences were acquired in all subjects on a 3.0T MRI system (Elition, Philips Healthcare, Best, the Netherlands). T1-weighted (T1w), T2-weighted (T2w) and Proton density-weighted imaging (PD) sequences were acquired with the musculoskeletal system as the main targets. The parameters used are shown in Table 1.These sequences were reconstructed by CS, CS-AI, and CS-AI-HR.
Image quality was evaluated by two radiologists with more than 5 years of experience. All images were rated according to an ordinal 5-point Likert scale (1 = poor, 2 = below average, 3 = fair, 4 = good, 5 = excellent), evaluating the following criteria: partial volume effect, blurring, discrimination from adjacent structures, and signal homogeneity. Friedman test with Dunn’s correction for multiple comparisons was used to compare the semi-quantitative image quality rates between CS, CS-AI and CS-AI-HR sequences. p-values <0.05 were considered significant.

Results

Examples of CS-AI-HR, CS-AI and CS images at lumbar region are shown in Figure 1 and Figure 2. In Figure 1, CS-AI-HR has a much clearer depiction of the hyperplasia and grossness of the vertebral body. In Figure 2, the presentation of an old compression fracture was better visible on the CS-AI-HR image compared to CS-AI or CS.. The image quality assessment of the two radiologists is provided in Table 2. The image quality of the routine scanned PDw images using CS-AI-HR was significantly higher than using CS (Figure 3, P = 0.0035).

Discussion

In our results, CS, CS-AI and CS-AI-HR images are all adequate for diagnosis without missed lesions. However, in terms of image clarity, CS-AI-HR enhanced the sharpness of the images, offers better image quality and is more descriptive for details.
While CS and CS-AI improve the SNR through the sparse representation of MR images, CS-AI-HR adds sharpening of image by its low- to high-resolution feature. This sharpening is present in all image features including artifacts that might exist e.g. due to motion. Even with the above limitation, CS-AI-HR was still scored better thanthan CS-AI or CS, which is in line with the finding from previous study².
The above results suggest that scan acceleration by CS-AI-HR reconstruction may help to obtain high quality images that can beimplemented in clinical routine.

Conclusion

CS-AI-HR reconstruction produce MR images with higher quality than that of CS images. It has potential to be integrated in clinical routine.

Acknowledgements

No acknowledgement found.