PA-GAN: Parallel Attention Based Gan for Enhancement of Fodf

Multi-shell HARDI scanning helps obtain a more accurate fiber orientation distribution function (fODF) than traditional DTI. However, it requires more DWI volumes, which require a longer scanning time, significantly limiting its clinical application. In this work, we have addressed this issue by proposing a GAN-based architecture to reconstruct Multi-shell Multi-tissue fODF (MSMT fODF), utilizing only a few DWI volumes obtained from single-shell scanning. The proposed GAN-based architecture leverages a Residual block, a Sigmoid-based Attention module, and a Feature inter-dependencies module. Besides, multiple loss functions are used, including total variation, L1 loss, and adversarial loss. Several qualitative and quantitative results justify our proposal in terms of MSMT fODF reconstruction, fiber tract estimation, and structural connectivity.