RETURN TO ABSTRACT LISTING


Objective:
To develop a deep learning-based model to generate Sampling Perfection with Application-optimized Contrasts using different flip-angle Evolution (SPACE) MRI pulse sequence - a less commonly acquired sequence - from magnetization-prepared rapid gradient echo (MPRAGE) images.

Materials and Methods:
We obtained 785 cases from our institutional imaging database, comprising brains with metastatic lesions. In 20 cases, both MP-RAGE and SPACE images were acquired simultaneously. For the rest of the patients, we found 189 cases with MP-RAGE available and 576 cases with SPACE images available. To generate SPACE images from MP-RAGE, we trained a cycle generative adversarial network (GAN). We performed inference on those 20 cases and generated SPACE from the MP-RAGE images, and compared the original SPACE with the generated SPACE images. The quality of the generated images was evaluated quantitatively by two neuroradiologists.

Results:
The figure shows an example case and the visual comparison between real MP-RAGE images, real SPACE images, and generated SPACE images. Upon qualitative evaluation, it was found that the generated SPACE images exhibited satisfactory visualization of contrast-enhanced lesions and suppressing the vessels. This study presents a proof of concept demonstrating the feasibility of generating SPACE images from MP-RAGE. In future investigations, we plan to evaluate the generated images with metastatic lesions and compare the number of lesions detected by neuroradiologists or with automated techniques.

Conclusion:
Contrast-enhanced T1-weighted (CT1W) MRI is a highly valuable tool for obtaining critical information about different brain lesions. Various pulse sequence protocols can be employed to obtain CT1W images, with MP-RAGE being the most commonly used. However, postgadolinium enhancement of MP-RAGE images is often suboptimal. An alternative protocol is the SPACE sequence, which offers superior contrast sensitivity and a better suppression of vascular flow signal[1,2]. This is especially important for detecting metastatic lesions found near brain vessels. It should be noted, however, that SPACE MRI acquisition requires more advanced MR scanners and is not currently feasible in all medical centers. Using deep learning-based generative models for generating MRI pulse sequences has proved beneficial in clinical tasks. In this study, we developed a deep-learning model to generate SPACE from MP-RAGE.