Mixture density knowledge distillation in super-resolution reconstruction of mri medical images

MRI medical image reconstruction frequently suffers from a smoothness bias, resulting in sub-optimal multi-valued mapping fitting. Mixture Density Networks (MDNs) offer a potential solution by modeling multi-valued functions via multiple components. However, numerical instability in MDNs undermines their performance. Moreover, the super-resolution task is inherently difficult due to its ill-posed nature.

To overcome these challenges, we introduce MixtUre densiTy knowlEdge Distillation (MUTED), a novel framework for super-resolution reconstruction. MUTED integrates the MDN module to mitigate boundary blurring, addresses MDN's numerical instability via an adversarial approach, and employs regularization derived from knowledge distillation to handle the ill-posed problem.

Extensive experiments on the IXI and BraTS21 datasets show that our MUTED framework effectively produces high-quality reconstructions. It outperforms existing methods in handling boundary blurring and numerical instability, as evidenced by experimental and visualization results.

MUTED surpasses state-of-the-art (SOTA) methods with a reduced computational cost and outperforms competing knowledge distillation methods. By addressing numerical instability and leveraging the regularization constraint, MUTED offers a robust solution for high-quality image reconstruction. Furthermore, the aleatoric uncertainty formulated by the MDN serves to reveal sharpened boundaries. This, in turn, effectively facilitates the efficient enhancement of the super-resolution reconstruction quality.