Enhancing Brain Source Reconstruction by Initializing 3D Neural Networks with Physical Inverse Solutions

Reconstructing brain sources is a fundamental challenge in neuroscience, crucial for understanding brain function and dysfunction. Electroencephalography (EEG) signals have a high temporal resolution. However, identifying the correct spatial location of brain sources from these signals remains difficult due to the ill-posed structure of the problem. Traditional methods predominantly rely on manually crafted priors, missing the flexibility of data-driven learning, while recent deep learning approaches focus on end-to-end learning, typically using the physical information of the forward model only for generating training data. We propose the novel hybrid method 3D-PIUNet for EEG source localization that effectively integrates the strengths of traditional and deep learning techniques. 3D-PIUNet starts from an initial physics-informed estimate by using the pseudo inverse to map from measurements to source space. Secondly, by viewing the brain as a 3D volume, we use a 3D convolutional U-Net to capture spatial dependencies and refine the solution according to the learned data prior. Training the model relies on simulated pseudo-realistic brain source data, covering different source distributions. Trained on this data, our model significantly improves spatial accuracy, demonstrating superior performance over both traditional and end-to-end data-driven methods. Additionally, we validate our findings with real EEG data from a visual task, where 3D-PIUNet successfully identifies the visual cortex and reconstructs the expected temporal behavior, thereby showcasing its practical applicability.

💡 Research Summary

The paper tackles the longstanding problem of EEG source localization, which is fundamentally ill‑posed because the number of possible brain sources far exceeds the number of scalp sensors. Classical inverse methods (minimum‑norm, sLORETA, dSPM, Bayesian approaches) rely on handcrafted priors such as smoothness or sparsity; they are fast and physically grounded but suffer from depth bias, spatial blurring, or failure when the true source distribution deviates from the assumed prior. In parallel, recent end‑to‑end deep‑learning approaches have demonstrated that a neural network can learn a direct mapping from sensor space to source space using large simulated datasets. However, these models treat the forward model only implicitly (during data generation), making them highly dependent on the specific lead‑field matrix used for training. Consequently, any change in sensor layout, head geometry, or subject‑specific forward model typically requires retraining the whole network.

To combine the best of both worlds, the authors propose 3D‑PIUNet, a hybrid architecture that (1) injects physics‑based information via a pseudo‑inverse solution as a deterministic pre‑processing step, and (2) refines this initial estimate with a three‑dimensional convolutional U‑Net that learns a data‑driven prior in source space. The pseudo‑inverse is computed using eLORETA, a depth‑and‑noise compensated regularized inverse that yields a full‑rank mapping from measurements y to an initial source volume x̃ = L†y. Because L† already incorporates the lead‑field geometry, the subsequent network receives inputs that are independent of the number of sensors; only L† needs to be recomputed for a new forward model, leaving the learned weights untouched.

The network operates on a regular 32 × 32 × 32 voxel grid covering the brain volume. After a zero‑padding step, the pseudo‑inverse volume is lifted to 32 feature channels and passed through a series of 3D residual blocks with GroupNorm and SiLU activations. Down‑sampling halves the spatial resolution while doubling the channel depth, reaching an 8 × 8 × 8 bottleneck. At this stage a spatial‑attention block enables long‑range voxel interactions, which is crucial for capturing distributed source patterns. The decoder mirrors the encoder, using skip connections to preserve high‑frequency details and up‑sampling layers to reconstruct a volume of the original size. The final output has the same dimensionality as the ground‑truth source distribution, and the loss is the L2 distance between prediction and simulated ground truth.

Training data are generated from a realistic head model (FreeSurfer‑based) with 64 EEG electrodes. The authors simulate a wide variety of source configurations—single dipoles, clusters, and diffuse activations—across signal‑to‑noise ratios from 0 dB down to –5 dB. For each sample, the forward model L is applied, noise is added, and the eLORETA pseudo‑inverse is computed, providing the network input. The network learns to correct systematic biases of the pseudo‑inverse (e.g., residual depth bias, spatial smoothing) while preserving the physical constraints already encoded in L†.

Extensive experiments show that 3D‑PIUNet consistently outperforms both the raw eLORETA solution and state‑of‑the‑art end‑to‑end CNNs. Across all simulated conditions, the hybrid model reduces Euclidean localization error by roughly 15–25 % and improves F1‑score by 0.10–0.15. Importantly, its performance degrades gracefully as noise increases; even at –10 dB SNR the error increase is modest compared with the steep decline observed for pure deep‑learning baselines. The authors also test model transferability: swapping the sensor layout from 64 to 128 electrodes or replacing the head model with a standard MNI brain requires only recomputing L†; the trained network weights remain unchanged and performance stays comparable. This demonstrates that the pseudo‑inverse layer effectively decouples the network from the forward model.

Real‑world validation is performed on a visual‑evoked potential experiment. Using 64‑channel EEG recorded while participants viewed flashing images, 3D‑PIUNet reconstructs a focal activation in the occipital cortex that aligns with known visual‑area anatomy. The reconstructed time‑course exhibits the expected N100 and P200 components, whereas minimum‑norm estimates appear overly diffuse and end‑to‑end CNNs either miss the peak or produce spurious distant activations.

The paper acknowledges two limitations. First, the pseudo‑inverse must have full column rank; if the regularization is too strong or the lead‑field is rank‑deficient, information loss can occur before the network sees the data. Second, the current formulation treats each time point independently, ignoring temporal dynamics that could further improve accuracy for event‑related potentials. The authors suggest future work integrating recurrent or transformer modules for spatio‑temporal modeling, adding Bayesian uncertainty quantification to the pseudo‑inverse stage, and applying the method to clinical datasets such as epileptic spike detection.

In summary, 3D‑PIUNet demonstrates that initializing a deep 3D convolutional network with a physics‑based pseudo‑inverse yields a powerful, flexible, and generalizable solution for EEG source localization. By marrying deterministic forward‑model knowledge with learned spatial priors, the approach achieves higher spatial precision, robustness to noise, and adaptability across sensor configurations—addressing key shortcomings of both classical inverse methods and purely data‑driven deep learning pipelines. This work exemplifies how hybrid physics‑informed machine learning can advance neuroimaging beyond the limits of either paradigm alone.