Abstract
Functional MRI (fMRI) at ultra-high fields (7T) offers unprecedented spatial resolution but is constrained by a fundamental trade-off between speed, coverage, and signal-to-noise ratio. Traditional acceleration techniques fail at extreme undersampling rates, while supervised deep learning is limited by the lack of fully-sampled "ground truth" data. We present a self-supervised deep learning reconstruction method that achieves 20-fold acceleration without requiring reference data. By employing a physics-guided unrolled network trained directly on undersampled k-space, we demonstrate image quality and functional signal integrity comparable to standard 10-fold acquisitions, significantly outperforming conventional parallel imaging methods.
Introduction
Mapping dynamic brain activity requires high temporal and spatial resolution. To achieve this at 7T, researchers utilize Simultaneous Multi-Slice (SMS) and in-plane undersampling. However, as acceleration factors increase toward 20-fold, traditional reconstruction algorithms like GRAPPA or SPSG (Sparse SENSE-GRAPPA) introduce severe aliasing artifacts and spatially varying "g-factor" noise.
Deep learning (DL) has emerged as a powerful alternative, but standard supervised models require fully-sampled scans for training. In high-resolution fMRI, such scans are often impossible to acquire due to the slow speed of full k-space traversal relative to neural dynamics. This work addresses this bottleneck using a referenceless, self-supervised approach.
Methods
The proposed method utilizes Self-Supervised learning via Data Undersampling (SSDU) integrated into a physics-guided unrolled network. Unlike "black box" models, this architecture alternates between a deep learning regularizer (ResNet) and a Data Consistency (DC) unit to ensure the output remains faithful to the physical acquisition parameters.
To train without ground truth, the acquired undersampled k-space data is split into two disjoint sets. One set is used within the DC units, while the other serves as the target for the loss function. This forces the network to learn to interpolate missing data points based on observed ones. For 20-fold acceleration, a multi-mask strategy ($K=6$) was used to provide diverse training signals, and slices were concatenated along the readout direction to exploit inter-slice correlations.
Experimental Setup
Data was acquired from 8 subjects on a 7T Siemens system using a 32-channel head coil. The protocol followed the Human Connectome Project (HCP) retinotopy paradigm (visual stimulation with rotating wedges). The model was trained on a single time-frame from 4 subjects and tested on the full time series of 4 unseen subjects to evaluate functional stability.
Results
The self-supervised DL approach successfully achieved a 20-fold acceleration (5x SMS and 4x in-plane). Qualitatively, the DL reconstruction effectively removed residual aliasing and ghosting artifacts that were prominent in the 20x SPSG reconstructions.
Quantitatively, the temporal Signal-to-Noise Ratio (tSNR) for the 20x DL reconstruction (~17.38) remained remarkably close to the 10x SPSG baseline (~18.99), whereas the 20x SPSG method suffered a catastrophic drop to ~8.94. Most importantly, functional coherence maps and phase maps remained intact, indicating that the network preserved the timing and integrity of the neural signals required for retinotopic mapping.
Discussion and Conclusion
This study demonstrates that self-supervised deep learning can push fMRI acceleration boundaries to 20x without the need for impossible-to-acquire reference data. By suppressing g-factor noise and aliasing while maintaining functional signal integrity, this method enables higher spatial resolution or faster temporal sampling for whole-brain 7T fMRI. Future work will focus on prospective 20x acquisitions and incorporating temporal redundancies to further enhance SNR.
Summary updated