Task-Dependent Warping of Semantic Representations During Search for Visual Action Categories.

Object and action perception in cluttered dynamic natural scenes relies on efficient allocation of limited brain resources to prioritize the attended targets over distractors. It has been suggested that during visual search for objects, distributed semantic representation of hundreds of object categories is warped to expand the representation of targets. Yet, little is known about whether and where in the brain visual search for action categories modulates semantic representations. To address this fundamental question, we studied brain activity recorded from five subjects (1 female) via functional magnetic resonance imaging while they viewed natural movies and searched for either communication or locomotion actions. We find that attention directed to action categories elicits tuning shifts that warp semantic representations broadly across neocortex, and that these shifts interact with intrinsic selectivity of cortical voxels for target actions. These results suggest that attention serves to facilitate task performance during social interactions by dynamically shifting semantic selectivity towards target actions, and that tuning shifts are a general feature of conceptual representations in the brain.SIGNIFICANCE STATEMENTThe ability to swiftly perceive the actions and intentions of others is a crucial skill for humans, which relies on efficient allocation of limited brain resources to prioritise the attended targets over distractors. However, little is known about the nature of high-level semantic representations during natural visual search for action categories. Here we provide the first evidence showing that attention significantly warps semantic representations by inducing tuning shifts in single cortical voxels, broadly spread across occipitotemporal, parietal, prefrontal, and cingulate cortices. This dynamic attentional mechanism can facilitate action perception by efficiently allocating neural resources to accentuate the representation of task-relevant action categories.

Attentional Modulation of Hierarchical Speech Representations in a Multitalker Environment.

Humans are remarkably adept in listening to a desired speaker in a crowded environment, while filtering out nontarget speakers in the background. Attention is key to solving this difficult cocktail-party task, yet a detailed characterization of attentional effects on speech representations is lacking. It remains unclear across what levels of speech features and how much attentional modulation occurs in each brain area during the cocktail-party task. To address these questions, we recorded whole-brain blood-oxygen-level-dependent (BOLD) responses while subjects either passively listened to single-speaker stories, or selectively attended to a male or a female speaker in temporally overlaid stories in separate experiments. Spectral, articulatory, and semantic models of the natural stories were constructed. Intrinsic selectivity profiles were identified via voxelwise models fit to passive listening responses. Attentional modulations were then quantified based on model predictions for attended and unattended stories in the cocktail-party task. We find that attention causes broad modulations at multiple levels of speech representations while growing stronger toward later stages of processing, and that unattended speech is represented up to the semantic level in parabelt auditory cortex. These results provide insights on attentional mechanisms that underlie the ability to selectively listen to a desired speaker in noisy multispeaker environments.

Biased competition in semantic representation during natural visual search.

Humans divide their attention among multiple visual targets in daily life, and visual search can get more difficult as the number of targets increases. The biased competition hypothesis (BC) has been put forth as an explanation for this phenomenon. BC suggests that brain responses during divided attention are a weighted linear combination of the responses during search for each target individually. This combination is assumed to be biased by the intrinsic selectivity of cortical regions. Yet, it is unknown whether attentional modulation of semantic representations are consistent with this hypothesis when viewing cluttered, dynamic natural scenes. Here, we investigated whether BC accounts for semantic representation during natural category-based visual search. Subjects viewed natural movies, and their whole-brain BOLD responses were recorded while they attended to "humans", "vehicles" (i.e. single-target attention tasks), or "both humans and vehicles" (i.e. divided attention) in separate runs. We computed a voxelwise linearity index to assess whether semantic representation during divided attention can be modeled as a weighted combination of representations during the two single-target attention tasks. We then examined the bias in weights of this linear combination across cortical ROIs. We find that semantic representations of both target and nontarget categories during divided attention are linear to a substantial degree, and that they are biased toward the preferred target in category-selective areas across ventral temporal cortex. Taken together, these results suggest that the biased competition hypothesis is a compelling account for attentional modulation of semantic representations.

Informed feature regularization in voxelwise modeling for naturalistic fMRI experiments.

Voxelwise modeling is a powerful framework to predict single-voxel functional selectivity for the stimulus features that exist in complex natural stimuli. Yet, because VM disregards potential correlations across stimulus features or neighboring voxels, it may yield suboptimal sensitivity in measuring functional selectivity in the presence of high levels of measurement noise. Here, we introduce a novel voxelwise modeling approach that simultaneously utilizes stimulus correlations in model features and response correlations among voxel neighborhoods. The proposed method performs feature and spatial regularization while still generating single-voxel response predictions. We demonstrated the performance of our approach on a functional magnetic resonance imaging dataset from a natural vision experiment. Compared to VM, the proposed method yields clear improvements in prediction performance, together with increased feature coherence and spatial coherence of voxelwise models. Overall, the proposed method can offer improved sensitivity in modeling of single voxels in naturalistic functional magnetic resonance imaging experiments.

A Transfer-Learning Approach for Accelerated MRI Using Deep Neural Networks.

PURPOSE: Neural networks have received recent interest for reconstruction of undersampled MR acquisitions. Ideally, network performance should be optimized by drawing the training and testing data from the same domain. In practice, however, large datasets comprising hundreds of subjects scanned under a common protocol are rare. The goal of this study is to introduce a transfer-learning approach to address the problem of data scarcity in training deep networks for accelerated MRI. METHODS: Neural networks were trained on thousands (upto 4 thousand) of samples from public datasets of either natural images or brain MR images. The networks were then fine-tuned using only tens of brain MR images in a distinct testing domain. Domain-transferred networks were compared to networks trained directly in the testing domain. Network performance was evaluated for varying acceleration factors (4-10), number of training samples (0.5-4k), and number of fine-tuning samples (0-100). RESULTS: The proposed approach achieves successful domain transfer between MR images acquired with different contrasts (T1 - and T2 -weighted images) and between natural and MR images (ImageNet and T1 - or T2 -weighted images). Networks obtained via transfer learning using only tens of images in the testing domain achieve nearly identical performance to networks trained directly in the testing domain using thousands (upto 4 thousand) of images. CONCLUSION: The proposed approach might facilitate the use of neural networks for MRI reconstruction without the need for collection of extensive imaging datasets.

Factorized sensitivity estimation for artifact suppression in phase-cycled bSSFP MRI.

OBJECTIVE: Balanced steady-state free precession (bSSFP) imaging suffers from banding artifacts in the presence of magnetic field inhomogeneity. The purpose of this study is to identify an efficient strategy to reconstruct banding-free bSSFP images from multi-coil multi-acquisition datasets. METHOD: Previous techniques either assume that a naïve coil-combination is performed a priori resulting in suboptimal artifact suppression, or that artifact suppression is performed for each coil separately at the expense of significant computational burden. Here we propose a tailored method that factorizes the estimation of coil and bSSFP sensitivity profiles for improved accuracy and/or speed. RESULTS: In vivo experiments show that the proposed method outperforms naïve coil-combination and coil-by-coil processing in terms of both reconstruction quality and time. CONCLUSION: The proposed method enables computationally efficient artifact suppression for phase-cycled bSSFP imaging with modern coil arrays. Rapid imaging applications can efficiently benefit from the improved robustness of bSSFP imaging against field inhomogeneity.

Simultaneous use of individual and joint regularization terms in compressive sensing: Joint reconstruction of multi-channel multi-contrast MRI acquisitions.

Multi-contrast images are commonly acquired together to maximize complementary diagnostic information, albeit at the expense of longer scan times. A time-efficient strategy to acquire high-quality multi-contrast images is to accelerate individual sequences and then reconstruct undersampled data with joint regularization terms that leverage common information across contrasts. However, these terms can cause features that are unique to a subset of contrasts to leak into the other contrasts. Such leakage-of-features may appear as artificial tissues, thereby misleading diagnosis. The goal of this study is to develop a compressive sensing method for multi-channel multi-contrast magnetic resonance imaging (MRI) that optimally utilizes shared information while preventing feature leakage. Joint regularization terms group sparsity and colour total variation are used to exploit common features across images while individual sparsity and total variation are also used to prevent leakage of distinct features across contrasts. The multi-channel multi-contrast reconstruction problem is solved via a fast algorithm based on Alternating Direction Method of Multipliers. The proposed method is compared against using only individual and only joint regularization terms in reconstruction. Comparisons were performed on single-channel simulated and multi-channel in-vivo datasets in terms of reconstruction quality and neuroradiologist reader scores. The proposed method demonstrates rapid convergence and improved image quality for both simulated and in-vivo datasets. Furthermore, while reconstructions that solely use joint regularization terms are prone to leakage-of-features, the proposed method reliably avoids leakage via simultaneous use of joint and individual terms, thereby holding great promise for clinical use.

Spatially informed voxelwise modeling for naturalistic fMRI experiments.

Voxelwise modeling (VM) is a powerful framework to predict single voxel responses evoked by a rich set of stimulus features present in complex natural stimuli. However, because VM disregards correlations across neighboring voxels, its sensitivity in detecting functional selectivity can be diminished in the presence of high levels of measurement noise. Here, we introduce spatially-informed voxelwise modeling (SPIN-VM) to take advantage of response correlations in spatial neighborhoods of voxels. To optimally utilize shared information, SPIN-VM performs regularization across spatial neighborhoods in addition to model features, while still generating single-voxel response predictions. We demonstrated the performance of SPIN-VM on a rich dataset from a natural vision experiment. Compared to VM, SPIN-VM yields higher prediction accuracies and better capture locally congruent information representations across cortex. These results suggest that SPIN-VM offers improved performance in predicting single-voxel responses and recovering coherent information representations.

Reconstruction by calibration over tensors for multi-coil multi-acquisition balanced SSFP imaging.

PURPOSE: To develop a rapid imaging framework for balanced steady-state free precession (bSSFP) that jointly reconstructs undersampled data (by a factor of R) across multiple coils (D) and multiple acquisitions (N). To devise a multi-acquisition coil compression technique for improved computational efficiency. METHODS: The bSSFP image for a given coil and acquisition is modeled to be modulated by a coil sensitivity and a bSSFP profile. The proposed reconstruction by calibration over tensors (ReCat) recovers missing data by tensor interpolation over the coil and acquisition dimensions. Coil compression is achieved using a new method based on multilinear singular value decomposition (MLCC). ReCat is compared with iterative self-consistent parallel imaging (SPIRiT) and profile encoding (PE-SSFP) reconstructions. RESULTS: Compared to parallel imaging or profile-encoding methods, ReCat attains sensitive depiction of high-spatial-frequency information even at higher R. In the brain, ReCat improves peak SNR (PSNR) by 1.1 ± 1.0 dB over SPIRiT and by 0.9 ± 0.3 dB over PE-SSFP (mean ± SD across subjects; average for N = 2-8, R = 8-16). Furthermore, reconstructions based on MLCC achieve 0.8 ± 0.6 dB higher PSNR compared to those based on geometric coil compression (GCC) (average for N = 2-8, R = 4-16). CONCLUSION: ReCat is a promising acceleration framework for banding-artifact-free bSSFP imaging with high image quality; and MLCC offers improved computational efficiency for tensor-based reconstructions. Magn Reson Med 79:2542-2554, 2018. © 2017 International Society for Magnetic Resonance in Medicine.

Functional Subdomains within Scene-Selective Cortex: Parahippocampal Place Area, Retrosplenial Complex, and Occipital Place Area.

Functional MRI studies suggest that at least three brain regions in human visual cortex-the parahippocampal place area (PPA), retrosplenial complex (RSC), and occipital place area (OPA; often called the transverse occipital sulcus)-represent large-scale information in natural scenes. Tuning of voxels within each region is often assumed to be functionally homogeneous. To test this assumption, we recorded blood oxygenation level-dependent responses during passive viewing of complex natural movies. We then used a voxelwise modeling framework to estimate voxelwise category tuning profiles within each scene-selective region. In all three regions, cluster analysis of the voxelwise tuning profiles reveals two functional subdomains that differ primarily in their responses to animals, man-made objects, social communication, and movement. Thus, the conventional functional definitions of the PPA, RSC, and OPA appear to be too coarse. One attractive hypothesis is that this consistent functional subdivision of scene-selective regions is a reflection of an underlying anatomical organization into two separate processing streams, one selectively biased toward static stimuli and one biased toward dynamic stimuli. SIGNIFICANCE STATEMENT: Visual scene perception is a critical ability to survive in the real world. It is therefore reasonable to assume that the human brain contains neural circuitry selective for visual scenes. Here we show that responses in three scene-selective areas-identified in previous studies-carry information about many object and action categories encountered in daily life. We identify two subregions in each area: one that is selective for categories of man-made objects, and another that is selective for vehicles and locomotion-related action categories that appear in dynamic scenes. This consistent functional subdivision may reflect an anatomical organization into two processing streams, one biased toward static stimuli and one biased toward dynamic stimuli.

Profile-encoding reconstruction for multiple-acquisition balanced steady-state free precession imaging.

PURPOSE: The scan-efficiency in multiple-acquisition balanced steady-state free precession imaging can be maintained by accelerating and reconstructing each phase-cycled acquisition individually, but this strategy ignores correlated structural information among acquisitions. Here, an improved acceleration framework is proposed that jointly processes undersampled data across N phase cycles. METHODS: Phase-cycled imaging is cast as a profile-encoding problem, modeling each image as an artifact-free image multiplied with a distinct balanced steady-state free precession profile. A profile-encoding reconstruction (PE-SSFP) is employed to recover missing data by enforcing joint sparsity and total-variation penalties across phase cycles. PE-SSFP is compared with individual compressed-sensing and parallel-imaging (ESPIRiT) reconstructions. RESULTS: In the brain and the knee, PE-SSFP yields improved image quality compared to individual compressed-sensing and other tested methods particularly for higher N values. On average, PE-SSFP improves peak SNR by 3.8 ± 3.0 dB (mean ± s.e. across N = 2-8) and structural similarity by 1.4 ± 1.2% over individual compressed-sensing, and peak SNR by 5.6 ± 0.7 dB and structural similarity by 7.1 ± 0.5% over ESPIRiT. CONCLUSION: PE-SSFP attains improved image quality and preservation of high-spatial-frequency information at high acceleration factors, compared to conventional reconstructions. PE-SSFP is a promising technique for scan-efficient balanced steady-state free precession imaging with improved reliability against field inhomogeneity. Magn Reson Med 78:1316-1329, 2017. © 2016 International Society for Magnetic Resonance in Medicine.

Spectrally selective imaging with wideband balanced steady-state free precession MRI.

PURPOSE: Unwanted, bright fat signals in balanced steady-state free precession sequences are commonly suppressed using spectral shaping. Here, a new spectral-shaping method is proposed to significantly improve the uniformity of stopband suppression without compromising the level of passband signals. METHODS: The proposed method combines binomial-pattern excitation pulses with a wideband balanced steady-state free precession sequence kernel. It thereby increases the frequency separation between the centers of pass and stopbands by π radians, enabling improved water-fat contrast. Simulations were performed to find the optimal flip angles and subpulse spacing for the binomial pulses that maximize contrast and signal efficiency. RESULTS: Comparisons with a conventional binomial balanced steady-state free precession sequence were performed in simulations as well as phantom and in vivo experiments at 1.5 T and 3 T. Enhanced fat suppression is demonstrated in vivo with an average improvement of 58% in blood-fat and 68% in muscle-fat contrast (P < 0.001, Wilcoxon signed-rank test). CONCLUSION: The proposed binomial wideband balanced steady-state free precession method is a promising candidate for spectrally selective imaging with enhanced reliability against field inhomogeneities.

Targeted vessel reconstruction in non-contrast-enhanced steady-state free precession angiography.

Image quality in non-contrast-enhanced (NCE) angiograms is often limited by scan time constraints. An effective solution is to undersample angiographic acquisitions and to recover vessel images with penalized reconstructions. However, conventional methods leverage penalty terms with uniform spatial weighting, which typically yield insufficient suppression of aliasing interference and suboptimal blood/background contrast. Here we propose a two-stage strategy where a tractographic segmentation is employed to auto-extract vasculature maps from undersampled data. These maps are then used to incur spatially adaptive sparsity penalties on vascular and background regions. In vivo steady-state free precession angiograms were acquired in the hand, lower leg and foot. Compared with regular non-adaptive compressed sensing (CS) reconstructions (CSlow ), the proposed strategy improves blood/background contrast by 71.3 ± 28.9% in the hand (mean ± s.d. across acceleration factors 1-8), 30.6 ± 11.3% in the lower leg and 28.1 ± 7.0% in the foot (signed-rank test, P < 0.05 at each acceleration). The proposed targeted reconstruction can relax trade-offs between image contrast, resolution and scan efficiency without compromising vessel depiction.

Enhanced phase-sensitive SSFP reconstruction for fat-water separation in phased-array acquisitions.

PURPOSE: To propose and assess a method to improve the reliability of phase-sensitive fat-water separation for phased-array balanced steady-state free precession (bSSFP) acquisitions. Phase-sensitive steady-state free precession (PS-SSFP) is an efficient fat-water separation technique that detects the phase difference between neighboring bands in the bSSFP magnetization profile. However, large spatial variations in the sensitivity profiles of phased-array coils can lead to noisy phase estimates away from the coil centers, compromising tissue classification. MATERIALS AND METHODS: We first perform region-growing phase correction in individual coil images via unsupervised selection of a fat-voxel seed near the peak of each coil's sensitivity profile. We then use an optimal linear combination of phase-corrected images to segregate fat and water signals. The proposed method was demonstrated on noncontrast-enhanced SSFP angiograms of the thigh, lower leg, and foot acquired at 1.5T using an 8-channel coil. Individual coil PS-SSFP with a common seed selection for all coils, individual coil PS-SSFP with coil-wise seed selection, PS-SSFP after coil combination, and IDEAL reconstructions were also performed. Water images reconstructed via PS-SSFP methods were compared in terms of the level of fat suppression and the similarity to reference IDEAL images (signed-rank test). RESULTS: While tissue misclassification was broadly evident across regular PS-SSFP images, the proposed method achieved significantly higher levels of fat suppression (P < 0.005) and increased similarity to reference IDEAL images (P < 0.005). CONCLUSION: The proposed method enhances fat-water separation in phased-array acquisitions by producing improved phase estimates across the imaging volume. J. Magn. Reson. Imaging 2016;44:148-157.

Functional subdomains within human FFA.

The fusiform face area (FFA) is a well-studied human brain region that shows strong activation for faces. In functional MRI studies, FFA is often assumed to be a homogeneous collection of voxels with similar visual tuning. To test this assumption, we used natural movies and a quantitative voxelwise modeling and decoding framework to estimate category tuning profiles for individual voxels within FFA. We find that the responses in most FFA voxels are strongly enhanced by faces, as reported in previous studies. However, we also find that responses of individual voxels are selectively enhanced or suppressed by a wide variety of other categories and that these broader tuning profiles differ across FFA voxels. Cluster analysis of category tuning profiles across voxels reveals three spatially segregated functional subdomains within FFA. These subdomains differ primarily in their responses for nonface categories, such as animals, vehicles, and communication verbs. Furthermore, this segregation does not depend on the statistical threshold used to define FFA from responses to functional localizers. These results suggest that voxels within FFA represent more diverse information about object and action categories than generally assumed.

Hadamard slice encoding for reduced-FOV diffusion-weighted imaging.

PURPOSE: To improve the clinical utility of diffusion-weighted imaging (DWI) by extending the slice coverage of a high-resolution reduced field-of-view technique. THEORY: Challenges in achieving high spatial resolution restrict the use of DWI in assessment of small structures such as the spinal cord. A reduced field-of-view method with 2D echo-planar radiofrequency (RF) excitation was recently proposed for high-resolution DWI. Here, a Hadamard slice-encoding scheme is proposed to double the slice coverage by exploiting the periodicity of the 2D echo-planar RF excitation profile. METHODS: A 2D echo-planar RF pulse and matching multiband refocusing RF pulses were designed using the Shinnar-Le Roux algorithm to reduce band interference, and variable-rate selective excitation to shorten the pulse durations. Hadamard-encoded images were resolved through a phase-preserving image reconstruction. The performance of the method was evaluated via simulations, phantom experiments, and in vivo high-resolution axial DWI of spinal cord. RESULTS: The proposed scheme successfully extends the slice coverage, while preserving the sharp excitation profile and the reliable fat suppression of the original method. For in vivo axial DWI of the spinal cord, an in-plane resolution of 0.7 × 0.7 mm(2) was achieved with 16 slices. CONCLUSION: The proposed Hadamard slice-encoding scheme doubles the slice coverage of the 2D echo-planar RF reduced field-of-view method without any scan-time penalty.

Three-dimensional magnetization-prepared imaging using a concentric cylinders trajectory.

PURPOSE: To develop new magnetization-prepared imaging schemes based on a three-dimensional (3D) concentric cylinders trajectory. METHODS: The 3D concentric cylinders trajectory, which is robust to off-resonance effects and timing delays while requiring fewer excitations than a comparable 3D Cartesian (3DFT) sequence, is used as the readout for magnetization-prepared sequences exploiting its inherently centric-ordered structure. Two applications: (i) T1 -weighted brain imaging with an inversion-recovery-prepared radiofrequency-spoiled gradient-echo (IR-SPGR) sequence, (ii) non-contrast-enhanced (NCE) peripheral angiography with a magnetization-prepared balanced steady-state free precession (bSSFP) sequence are presented to demonstrate the effectiveness of the proposed method. For peripheral angiography, the scan efficiency is further improved by interleaving different preparations at different rates and by carefully designing the sampling geometry for an efficient parallel imaging method. RESULTS: In vivo brain scans with an IR-SPGR sequence and lower extremity scans with a magnetization-prepared bSSFP sequence for NCE peripheral angiography both demonstrate that the proposed sequences with concentric cylinders effectively capture the transient magnetization-prepared contrast with faster scan times than a corresponding 3DFT sequence. The application of peripheral angiography also shows the feasibility of the proposed interleaving schemes and parallel imaging method. CONCLUSION: The 3D concentric cylinders trajectory is a robust and efficient readout that is well-suited for magnetization-prepared imaging.

A Plug-In Graph Neural Network to Boost Temporal Sensitivity in fMRI Analysis.

Learning-based methods offer performance leaps over traditional methods in classification analysis of high-dimensional functional MRI (fMRI) data. In this domain, deep-learning models that analyze functional connectivity (FC) features among brain regions have been particularly promising. However, many existing models receive as input temporally static FC features that summarize inter-regional interactions across an entire scan, reducing the temporal sensitivity of classifiers by limiting their ability to leverage information on dynamic FC features of brain activity. To improve the performance of baseline classification models without compromising efficiency, here we propose a novel plug-in based on a graph neural network, GraphCorr, to provide enhanced input features to baseline models. The proposed plug-in computes a set of latent FC features with enhanced temporal information while maintaining comparable dimensionality to static features. Taking brain regions as nodes and blood-oxygen-level-dependent (BOLD) signals as node inputs, GraphCorr leverages a node embedder module based on a transformer encoder to capture dynamic latent representations of BOLD signals. GraphCorr also leverages a lag filter module to account for delayed interactions across nodes by learning correlational features of windowed BOLD signals across time delays. These two feature groups are then fused via a message passing algorithm executed on the formulated graph. Comprehensive demonstrations on three public datasets indicate improved classification performance for several state-of-the-art graph and convolutional baseline models when they are augmented with GraphCorr.

COVID-19 Detection From Respiratory Sounds With Hierarchical Spectrogram Transformers.

Monitoring of prevalent airborne diseases such as COVID-19 characteristically involves respiratory assessments. While auscultation is a mainstream method for preliminary screening of disease symptoms, its utility is hampered by the need for dedicated hospital visits. Remote monitoring based on recordings of respiratory sounds on portable devices is a promising alternative, which can assist in early assessment of COVID-19 that primarily affects the lower respiratory tract. In this study, we introduce a novel deep learning approach to distinguish patients with COVID-19 from healthy controls given audio recordings of cough or breathing sounds. The proposed approach leverages a novel hierarchical spectrogram transformer (HST) on spectrogram representations of respiratory sounds. HST embodies self-attention mechanisms over local windows in spectrograms, and window size is progressively grown over model stages to capture local to global context. HST is compared against state-of-the-art conventional and deep-learning baselines. Demonstrations on crowd-sourced multi-national datasets indicate that HST outperforms competing methods, achieving over 90% area under the receiver operating characteristic curve (AUC) in detecting COVID-19 cases.

DEQ-MPI: A Deep Equilibrium Reconstruction With Learned Consistency for Magnetic Particle Imaging.

Magnetic particle imaging (MPI) offers unparalleled contrast and resolution for tracing magnetic nanoparticles. A common imaging procedure calibrates a system matrix (SM) that is used to reconstruct data from subsequent scans. The ill-posed reconstruction problem can be solved by simultaneously enforcing data consistency based on the SM and regularizing the solution based on an image prior. Traditional hand-crafted priors cannot capture the complex attributes of MPI images, whereas recent MPI methods based on learned priors can suffer from extensive inference times or limited generalization performance. Here, we introduce a novel physics-driven method for MPI reconstruction based on a deep equilibrium model with learned data consistency (DEQ-MPI). DEQ-MPI reconstructs images by augmenting neural networks into an iterative optimization, as inspired by unrolling methods in deep learning. Yet, conventional unrolling methods are computationally restricted to few iterations resulting in non-convergent solutions, and they use hand-crafted consistency measures that can yield suboptimal capture of the data distribution. DEQ-MPI instead trains an implicit mapping to maximize the quality of a convergent solution, and it incorporates a learned consistency measure to better account for the data distribution. Demonstrations on simulated and experimental data indicate that DEQ-MPI achieves superior image quality and competitive inference time to state-of-the-art MPI reconstruction methods.