⚡ electrical engineering

Coordinate-Queryable Neural Field Reconstruction for EEG Spatial Super-Resolution with Unseen-Electrode Generation

This paper proposes a coordinate-queryable neural field framework that reformulates EEG spatial super-resolution as learning a shared conditional scalp field, enabling robust reconstruction of both random missing channels and strictly unseen electrode signals by encoding observed channels with their coordinates into a latent condition for coordinate-based signal synthesis.

Original authors: Hongjun Liu, Leyu Zhou, Zijianghao Yang, Chao Yao

Published 2026-06-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hongjun Liu, Leyu Zhou, Zijianghao Yang, Chao Yao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is a bustling city, and EEG (electroencephalography) is a team of reporters trying to interview the citizens to understand what's happening. Usually, you want as many reporters as possible (high-density electrodes) to get a perfect picture. But in the real world, things go wrong: some reporters might be sick (bad contacts), some might be missing entirely (random channel loss), or you might only have a few reporters available to save money or time (low-density setup).

Most current methods for fixing this are like rigid translation dictionaries. If you give them a specific set of missing words (channels), they can translate them. But if the missing words change, or if you ask them to translate a word they've never seen before, they get confused and fail.

This paper introduces a new approach called ScalpINR. Instead of a rigid dictionary, think of it as a master mapmaker who learns the "terrain" of the brain's electrical activity.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Fixed Layout" Trap

Imagine you have a puzzle with 64 pieces. Old methods learn how to fill in the missing pieces only if you tell them exactly which pieces are missing (e.g., "Pieces 5, 12, and 40 are gone"). If you suddenly say, "Actually, pieces 5, 12, and 99 are gone," the old method panics because it was never trained on that specific combination. It tries to guess based on a fixed pattern rather than understanding the actual shape of the puzzle.

2. The Solution: A "Coordinate-Queryable" Map

ScalpINR changes the game. Instead of learning to fill in specific puzzle slots, it learns to draw a continuous, smooth map of the brain's electrical field.

The Encoder (The Observer): When the system sees some working reporters (visible electrodes), it doesn't just look at their numbers. It looks at where they are standing (their 3D coordinates) and what they are saying. It combines this into a "summary" of the current situation.
The Decoder (The Mapmaker): This is the magic part. Once the system has the summary, you can ask it: "What would the signal be at this specific coordinate?" It doesn't matter if that coordinate was one of the reporters you saw, or a spot in the middle of nowhere where no reporter ever stood. The system queries its internal map and generates the answer based on the geometry of the brain, not a pre-set list.

3. The Training Trick: "The Corrupted Reporter"

To make sure this mapmaker doesn't start making things up, the authors used a clever training strategy called Fidelity-Preserving Channel Corruption.

Imagine you are teaching an artist to draw a landscape.

Standard training: You show them a photo with some trees missing and ask them to draw the missing trees.
ScalpINR training: You show them the photo, but you also smudge some of the trees that are there (corrupted electrodes) and ask them to redraw those smudged trees correctly, while also filling in the missing ones.

This forces the artist to pay close attention to the actual details of the photo they are holding, ensuring the new drawing stays true to the original evidence, rather than drifting into fantasy.

4. The Results: Why It Matters

The paper tested this on three different brain datasets (emotions, attention, and motor imagery). Here is what they found:

Handling the Unknown: When the system was asked to generate signals for electrode locations it had never seen during training (strict "unseen-electrode" generation), it crushed the competition. On one dataset, it reduced errors by 37.5% compared to the best existing method.
Surviving Sparse Data: Even when only a tiny fraction of electrodes were working (as low as 6% or 12%), ScalpINR could still reconstruct the full brain activity much better than older methods.
Speed: Unlike some complex AI models that take a long time to "think" (like diffusion models), ScalpINR is fast. It can generate a full brain signal in about 2.6 milliseconds, making it practical for real-time use.

The Bottom Line

Think of old methods as a stencil: they only work if you hold the stencil in the exact right position. ScalpINR is like a flexible, living mold. It understands the shape of the brain's electrical field so well that it can fill in gaps or create new points anywhere on the surface, as long as it has a few reference points to guide it. This makes it incredibly robust for real-world scenarios where equipment might fail or where you need to use fewer sensors than usual.

Technical Summary: ScalpINR – Coordinate-Queryable Neural Field Reconstruction for EEG Spatial Super-Resolution

1. Problem Statement

Electroencephalography (EEG) spatial super-resolution (EEGSR) aims to reconstruct high-density (HD) EEG signals from sparse, low-density (LD) observations. While critical for practical applications where HD acquisition is hindered by cost, preparation burden, or hardware limitations, existing methods face significant robustness challenges in real-world deployments.

Current state-of-the-art approaches typically learn a fixed mapping from a predefined low-density input layout to a fixed high-density output layout. This formulation assumes a static set of input and output electrodes. However, in realistic scenarios, electrode quality fluctuates due to bad contacts, motion artifacts, or device variability, leading to random channel missingness and changing visible-channel patterns. When the set of missing channels varies at test time, fixed-mapping models often fail because they rely on channel-index-specific correlations rather than learning a reconstruction rule grounded in scalp geometry. Furthermore, classical interpolation methods, while geometry-aware, tend to oversmooth fine-grained spatial structures and cannot generate signals at electrode locations that were never observed during training.

The paper posits that EEGSR under random missingness should not be treated as restoring a fixed output tensor from a masked input tensor. Instead, it should be reformulated as reconstructing a coordinate-queryable scalp field from partially observed support channels.

2. Methodology: ScalpINR

The authors propose ScalpINR (Support-Conditioned Implicit Neural Representation), a framework that learns a shared conditional scalp field from observed channels and reconstructs signals by querying arbitrary electrode coordinates. The architecture consists of three core components:

2.1 Position-Guided Channel Encoder

To handle variable input layouts, ScalpINR explicitly encodes the spatial configuration of available channels.

Tokenization: For each visible channel $c$ , the signal $x_c$ is combined with its 3D electrode coordinate $p_c$ . The coordinate is transformed via a position embedding function $\phi(\cdot)$ to create a spatial descriptor.
Aggregation: Signal representations and position embeddings are fused into channel-wise tokens, which are then aggregated by a permutation-invariant encoder (e.g., Set Transformer or similar) into a single latent scalp representation $z$ . This allows the model to summarize signal context and electrode layout regardless of which specific channels are visible.

2.2 Coordinate-Queryable Conditional INR Decoder

The decoder utilizes an Implicit Neural Representation (INR) to synthesize signals.

Query Mechanism: Given the latent condition $z$ , the decoder predicts the EEG value at any arbitrary query coordinate $p_q$ and time $t$ using a conditional neural field function: $\hat{x}(p_q, t) = f_\omega(z, \phi(p_q), \tau(t))$ .
Decoupling: This design decouples the encoding of observed data from the generation of target signals. The model is not restricted to pre-defined output slots; it can generate signals at "unseen" electrode locations simply by querying their coordinates.

2.3 Fidelity-Preserving Channel Corruption Training

To prevent the generated field from drifting away from observed evidence, especially under severe missingness, the authors introduce a specialized training strategy involving mixed electrode states:

Missing Electrodes: Signals are removed from the input to train recovery capabilities.
Corrupted Electrodes: Signals are perturbed to mimic unstable or degraded observations, improving robustness.
Visible-but-Predicted Electrodes: Channels remain in the input but are included in the prediction target. This prevents the model from treating observed channels as unconstrained pass-through signals.
Objective: The training loss combines a reconstruction loss on target channels ( $L_{rec}$ ) and a fidelity loss on visible-but-predicted channels ( $L_{fid}$ ), ensuring the inferred scalp field remains consistent with observed support.

3. Key Contributions

Problem Reformulation: The paper reframes EEGSR under random channel missingness as a support-conditioned, coordinate-queryable scalp reconstruction problem. This unifies variable missing-channel reconstruction and strict unseen-electrode signal generation within a single framework.
ScalpINR Architecture: The proposal of a position-guided implicit neural representation framework that encodes partial observations into a latent condition and reconstructs signals via coordinate querying, enabling signal generation at locations never seen during training.
Training Strategy: The introduction of a fidelity-preserving channel corruption strategy that trains the model under mixed visible, corrupted, and missing states, enforcing consistency between the inferred field and observed data.

4. Experimental Results

The framework was evaluated on three public datasets: SEED (emotion), AAD (auditory attention), and BCI2000 (motor imagery).

4.1 Random Missing-Channel Reconstruction

Under varying support ratios ( $r$ ), ScalpINR consistently outperformed baselines (including Spherical Spline, ZUNA, ImputeINR, CSDI, ESTformer, MCMA, and SRGDiff) in signal-level fidelity (NMSE, PCC, SNR) and spectral preservation.

Performance: On the SEED dataset with a sparse support ratio of $r=0.125$ , ScalpINR reduced NMSE by 0.10 compared to the strongest baseline (MCMA) and improved SNR by 0.52 dB.
Spectral Fidelity: The method achieved the lowest spectral error across all datasets, indicating superior preservation of EEG frequency-domain structure compared to point-wise waveform similarity.
Downstream Utility: Reconstructed signals maintained task-discriminative information, yielding the best or tied-best accuracy in downstream decoding tasks (e.g., emotion recognition, motor imagery).

4.2 Strict Unseen-Electrode Generation

In a stricter protocol where a subset of electrodes was held out entirely during training (never used as input), ScalpINR demonstrated the ability to synthesize signals at these novel locations.

AAD Results: Under a strict held-out setting with full support ( $r=1$ for visible channels), ScalpINR reduced NMSE by 37.5% and improved SNR by 2.12 dB over the strongest baseline.
Generalization: The model successfully extrapolated to unseen electrode locations, whereas interpolation-based methods tended to over-smooth details, and other deep learning baselines showed significant deviations from ground truth.

4.3 Efficiency

ScalpINR is computationally efficient, requiring only 2.563M parameters and achieving an inference latency of 2.591 ms per sample. This is significantly faster than diffusion-based baselines (e.g., SRGDiff at ~92 ms) while maintaining higher accuracy.

5. Significance and Claims

The paper claims that ScalpINR addresses the brittleness of existing EEGSR methods in real-world deployments where channel availability is unpredictable. By shifting from fixed-mapping learning to coordinate-queryable neural field reconstruction, the method achieves:

Robustness: The ability to handle random channel missingness and varying electrode layouts without retraining or architectural changes.
Extrapolation: The capability to synthesize signals at electrode locations that were never exposed as input during training, a feat difficult for standard interpolation or fixed-mapping deep learning models.
Fidelity: The generation of signals that preserve both temporal morphology and spectral structure, ensuring utility for downstream decoding tasks.

The authors conclude that this coordinate-queryable paradigm offers a flexible and robust solution for EEG acquisition under variable channel availability, with future work planned to extend this to cross-device montage transfer and heterogeneous real-world conditions.