LFANET: Transforming 3T Single-Shell to 7T Multi-Shell DMRI Using Deep Learning Based Leapfrog and Attention

Ranjeet Ranjan Jha, Sudhir K. Pathak, Walter Schneider, B. V. Rathish Kumar, Arnav Bhavsar, Aditya Nigam

Abstract: HARDI-based diffusion MRI acquisition technique is a relatively recent modality of interest as it can yield more accurate fiber tracts. Besides, HARDI at higher magnetic strength is more sensitive to tissue changes and accurately estimate anatomical details in the human brain. However, a higher magnetic strength scanner is costly and not available in most clinical settings. Furthermore, due to signal-to-noise ratio issues and severe imaging artefacts, most existing 3T dMRI scanners with low gradient-strengths generally acquire single-shell up to b = 1000s/mm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> . Hence, in this work, we consider the task of transforming the 3T single-shell HARDI signal (at b = 1000s/mm <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> ) to a 7T multi-shell HARDI signal utilizing the proposed deep learning model LF ANet. The proposed model consists of modules based on a Leapfrog method and an attention module. In addition, we have included suitable loss functions such as L <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> and total variation loss. Several quantitative and qualitative results have been presented to show the effectiveness of the proposed method.