Learning Sampling and Model-Based Signal Recovery for Compressed Sensing MRI
This work addresses the need for faster and higher-quality MRI acquisitions, representing an incremental improvement in domain-specific methods.
The paper tackles the problem of designing optimal k-space sampling patterns for compressed sensing MRI by jointly learning task-adaptive sampling and a model-based reconstruction network, resulting in improved MR image quality compared to baseline methods.
Compressed sensing (CS) MRI relies on adequate undersampling of the k-space to accelerate the acquisition without compromising image quality. Consequently, the design of optimal sampling patterns for these k-space coefficients has received significant attention, with many CS MRI methods exploiting variable-density probability distributions. Realizing that an optimal sampling pattern may depend on the downstream task (e.g. image reconstruction, segmentation, or classification), we here propose joint learning of both task-adaptive k-space sampling and a subsequent model-based proximal-gradient recovery network. The former is enabled through a probabilistic generative model that leverages the Gumbel-softmax relaxation to sample across trainable beliefs while maintaining differentiability. The proposed combination of a highly flexible sampling model and a model-based (sampling-adaptive) image reconstruction network facilitates exploration and efficient training, yielding improved MR image quality compared to other sampling baselines.