Neuromorphic X-ray Computed Tomography¶

Conference: ECCV 2026
Paper: ECCV Official
Code: https://wanghongjian98.github.io/projects/neuroxct/
Area: Medical Imaging
Keywords: Neuromorphic Sensing, X-ray Computed Tomography, Implicit Neural Representations, Sparse-view Reconstruction, Event Camera

TL;DR¶

Presents the first application of neuromorphic sensing to X-ray CT and proposes the Event-enhanced Neural Attenuation Field (ENAF), which couples sparse intensity projections with microsecond asynchronous event streams to resolve angular ambiguities and significantly boost sparse-view reconstruction accuracy and efficiency.

Background & Motivation¶

X-ray computed tomography (CT) enables non-destructive volumetric imaging of internal morphology and serves as a foundational tool across medical diagnostics, materials science, and industrial non-destructive testing. Producing high-fidelity volumetric reconstructions conventionally requires a dense set of angular projections captured over sufficient exposure durations. However, strict radiation dose constraints in clinical diagnostics and ultra-fast scan requirements in synchrotron facilities inherently restrict angular measurements. Under these conditions, the acquisition enters an extremely sparse-view regime, rendering the inverse reconstruction problem severely ill-posed.

Current sparse-view CT techniques primarily leverage model-based iterative reconstruction (e.g., SART) or coordinate-based implicit neural representations (INRs, such as NAF and SAX-NeRF). While continuous neural representations deliver noticeable improvements over classical analytic algorithms (FDK/FBP), their performance remains strictly bounded by Nyquist sampling constraints at the physical detector level. Conventional frame-based detectors expose at uniform temporal intervals, causing severe angular aliasing during rapid variations and accumulating redundant measurements in static or slow-changing regions. Post-hoc regularizations cannot easily restore information lost during physical measurement.

In contrast, neuromorphic event cameras operate asynchronously at the individual pixel level, triggering an event only when the change in log-intensity surpasses a contrast threshold. This sensing mechanism offers microsecond temporal resolution, exceptional dynamic range exceeding 120 dB, and minimal temporal redundancy. During continuous object rotation in CT scanning, transient changes in transmitted attenuation produce continuous streams of informative events between sparsely acquired projection frames. Core idea: introduce neuromorphic event sensing into X-ray CT via an Event-enhanced Neural Attenuation Field (ENAF), coupling the absolute attenuation baseline from sparse frames with normalized high-temporal-resolution event accumulation maps over arbitrary angular windows, eliminating angular ambiguity directly from the physical measurement level.

Method¶

Overall Architecture¶

ENAF takes two complementary modalities as input: a set of sparse attenuation projections captured at uniform angular intervals, and a continuous stream of asynchronous events recorded with microsecond temporal resolution throughout the rotation scan. The target output is a continuous 3D volumetric attenuation field $f_\Phi(\mathbf{x})$.

The overall pipeline consists of coordinate parameterization, forward line-integral projection rendering, and dual-modality consistency optimization. The 3D attenuation volume is parameterized using multi-resolution hash encodings coupled with a compact MLP. For any arbitrary projection angle, spatial sample points along penetrating X-ray paths are numerically integrated using Beer–Lambert's law to synthesize predicted projections. During training, the framework optimizes a joint objective: uniform projection angles are supervised against ground-truth frames to lock in absolute attenuation scales, while pairs of timestamps $[t_a, t_b]$ are sampled at random to supervise predicted projection differences against accumulated event maps via normalized pattern matching.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["3D Ray Sample Points x"] --> B["Multi-resolution Hash Neural Field<br/>Hash Encoding + Compact MLP"]
    B --> C["Numerical Ray-casting Integration<br/>Render Projections & Differences"]
    C --> D["Dual-branch Supervision"]
    D -->|Uniform Projection Angles| E["Projection Loss L_proj<br/>Anchors Global Attenuation Scale"]
    D -->|Random Intervals ta to tb| F["Normalized Event Loss L_event<br/>Aligns High-frequency Angular Change"]

Key Designs¶

1. Multi-resolution Hash Neural Attenuation Field: Efficient Coordinate Parameterization

Direct voxel grid representations in sparse-view settings suffer from severe overfitting artifacts, whereas deep coordinate MLPs exhibit slow convergence. ENAF adopts multi-resolution hash encodings combined with a shallow MLP to parameterize the continuous 3D attenuation field $f_\Phi(\mathbf{x}) = \text{MLP}(\gamma(\mathbf{x}); \Phi)$. For an X-ray ray $L_{\theta, x, y}$ originating at $\mathbf{o}$ with unit direction $\mathbf{d}$ at rotation angle $\theta$, the predicted line integral projection $\hat{A}_\theta(x, y)$ is evaluated through discrete numerical integration over $S$ sampled points: $$ \hat{A}\theta(x, y) = \int)\,\Delta s $$ This parameterization delivers high spatial representation fidelity while keeping GPU memory and ray rendering overhead minimal, enabling efficient back-propagation during continuous multi-angle sampling.}} f_\Phi(\mathbf{x})\,ds \approx \sum_{k=1}^S f_\Phi(\mathbf{o} + s_k \mathbf{d

2. Asynchronous Event Interval Accumulation: Physics-based Log-attenuation Mapping

Event cameras trigger per-pixel binary polarities $p_k \in \{-1, +1\}$ when the temporal change in log-intensity reaches a contrast threshold $C$. According to the Beer–Lambert law, measured intensity is $I_\theta(x, y) = I_0 \exp(-A_\theta(x, y))$, meaning the change in log-intensity relates directly to the inverted difference in line-integral attenuation: $\log I_{\theta_b} - \log I_{\theta_a} = A_{\theta_a} - A_{\theta_b}$. Rather than attempting fragile single-event alignment, ENAF samples angular intervals $[t_a, t_b]$ (corresponding to rotation angles $\theta_a$ and $\theta_b$) and accumulates event polarities for each pixel into a 2D event map: $$ E_{a,b}(x, y) = C \sum_{e_k \in \mathcal{E}_{a,b}(x, y)} p_k $$ Aligning the synthesized projection difference $\Delta \hat{A}_{a,b} = \hat{A}_{\theta_a}(x, y) - \hat{A}_{\theta_b}(x, y)$ with $E_{a,b}(x, y)$ allows ENAF to accurately enforce intermediate angular dynamics between sparsely spaced frames.

3. Scale-decoupled Normalized Spatial Pattern Loss: Bypassing Threshold Calibration

In real hardware setups, analog event sensor properties, thermal drift, and pixel mismatch cause variations in the effective threshold $C$ across space and time. Directly computing an $L_2$ loss on unnormalized attenuation differences would propagate contrast calibration errors into the reconstructed volume. ENAF overcomes this by applying an $L_2$-normalized mean squared error loss: $$ \mathcal{L}{\mathrm{event}}(\theta_a, \theta_b) = \left| \frac{\Delta\hat{A}}}{|\Delta\hat{A{a,b}|_2 + \epsilon} - \frac{E \right|_2^2 $$ By decoupling absolute amplitude and matching normalized spatial gradient patterns, ENAF remains fully resilient to hardware contrast threshold calibration mismatches.}}{|E_{a,b}|_2 + \epsilon

Loss & Training¶

The parameters $\Phi$ are optimized end-to-end using a joint objective function: $$ \mathcal{L}(\Phi) = \mathbb{E}{\theta_n \sim \mathcal{P}\Theta}[\mathcal{L}{\mathrm{proj}}(\theta_n)] + \lambda_e \mathbb{E}E}[\mathcal{L}(\theta_a, \theta_b)] $$ Here, }$\mathcal{L}_{\mathrm{proj}}(\theta_n) = \|\hat{A}_{\theta_n} - A_{\theta_n}\|_2^2$ ensures global attenuation scale fidelity across sparse views, while $\mathcal{L}_{\mathrm{event}}$ provides dense angular gradient supervision. Optimization is handled using the Adam optimizer with a batch size of 1,024 rays for 3,000 epochs. The learning rate is set to $1 \times 10^{-3}$ and decays to $1 \times 10^{-4}$ over the first 1,500 epochs. Training is conducted on a single NVIDIA H100 GPU.

Key Experimental Results¶

Main Results¶

The method is systematically evaluated on four simulated anatomical benchmarks (Chest, Abdomen, Foot, Jaw) and one physical synchrotron beamline dataset. The quantitative comparison (PSNR [dB] / SSIM) across 5, 10, 25, and 50 sparse views is detailed below:

Dataset	Views	FDK	SART	NAF	SAX-NeRF	ENAF (Ours)	Gain (vs Prev. SOTA)
Chest	5 views	12.59 / 0.2082	19.83 / 0.5787	17.75 / 0.5157	20.52 / 0.6498	29.79 / 0.9330	+9.27 dB / +0.2832
	10 views	16.66 / 0.3651	23.42 / 0.7347	24.75 / 0.7965	25.53 / 0.8240	31.10 / 0.9479	+5.57 dB / +0.1239
	25 views	21.82 / 0.6208	28.15 / 0.8955	29.62 / 0.9168	30.46 / 0.9291	32.58 / 0.9622	+2.12 dB / +0.0331
	50 views	22.89 / 0.7861	32.17 / 0.9594	33.50 / 0.9667	34.26 / 0.9701	34.57 / 0.9750	+0.31 dB / +0.0049
Abdomen	5 views	12.19 / 0.1674	23.10 / 0.7246	21.42 / 0.6886	24.10 / 0.8005	29.14 / 0.8851	+5.04 dB / +0.0846
	10 views	15.62 / 0.2797	25.12 / 0.7885	24.80 / 0.7690	26.62 / 0.8462	30.23 / 0.9038	+3.61 dB / +0.0576
	25 views	19.20 / 0.4623	28.26 / 0.8598	29.53 / 0.8671	31.06 / 0.9192	32.76 / 0.9368	+1.70 dB / +0.0176
	50 views	22.63 / 0.6030	31.40 / 0.9170	34.69 / 0.9544	34.95 / 0.9589	35.11 / 0.9584	+0.16 dB / -0.0005
Foot	5 views	11.79 / 0.1593	22.67 / 0.7100	24.29 / 0.8357	23.65 / 0.8326	30.02 / 0.9105	+5.73 dB / +0.0748
	10 views	15.20 / 0.2371	24.34 / 0.7930	26.59 / 0.8703	26.74 / 0.8737	30.78 / 0.9223	+4.04 dB / +0.0486
	25 views	20.22 / 0.4141	27.26 / 0.8768	30.14 / 0.9135	30.32 / 0.9170	31.70 / 0.9330	+1.38 dB / +0.0160
	50 views	24.53 / 0.6000	30.29 / 0.9296	31.82 / 0.9347	32.25 / 0.9404	32.45 / 0.9409	+0.20 dB / +0.0005
Jaw	5 views	15.49 / 0.1288	24.80 / 0.7091	24.99 / 0.6869	25.12 / 0.7017	31.88 / 0.9026	+6.76 dB / +0.1935
	10 views	18.68 / 0.2496	26.04 / 0.7428	26.87 / 0.7581	27.41 / 0.7764	33.16 / 0.9267	+5.75 dB / +0.1503
	25 views	23.71 / 0.5386	29.07 / 0.8447	30.37 / 0.8629	31.39 / 0.8922	34.66 / 0.9467	+3.27 dB / +0.0545
	50 views	28.58 / 0.7816	33.13 / 0.9301	34.37 / 0.9398	35.34 / 0.9510	35.54 / 0.9530	+0.20 dB / +0.0020

On the real-world nano-architected metallic lattice dataset collected at the TOMCAT beamline of the Swiss Light Source, ENAF reconstructs sharp node structures and clear rod connectivity using as few as 5 frame projections combined with event streams, whereas frame-only baselines produce overwhelming streak artifacts.

Ablation Study¶

The noise robustness of ENAF under timestamp jitter and spurious noise is validated on the Chest dataset using 5 views:

Noise Type	Intensity Setting	PSNR (dB)	SSIM	Note
Clean Baseline	Clean	29.79	0.9330	Ideal simulated event stream
Timestamp Jitter ($t_{\max}$)	Low (1.0 ms)	29.78	0.9328	Negligible 0.01 dB drop under millisecond drift
	Medium (5.0 ms)	29.66	0.9312	Highly robust under 5 ms temporal delay
	High (10.0 ms)	29.49	0.9294	Drops only 0.30 dB even under 10 ms jitter
Spurious Noise (Ratio)	Low (10%)	29.77	0.9322	10% isolated false events added
	Medium (20%)	29.73	0.9316	20% random noise injected, drop is only 0.06 dB
	High (50%)	29.62	0.9292	With 50% extra noise, PSNR decreases by just 0.17 dB

An event-only reconstruction ablation further demonstrates that using only $\mathcal{L}_{\mathrm{event}}$ recovers distinct internal geometric boundaries, but lacks absolute absorption scales. Adding sparse frame projections anchors the quantitative values, proving the necessity of hybrid fusion.

Key Findings¶

Massive gains under extreme sparsity: In the most constrained 5-view scenario, ENAF outperforms SAX-NeRF by up to 9.27 dB and standard NAF by 12.04 dB on Chest, verifying that event streams supply critical high-frequency angular variation missing between frames.
Intrinsic noise resilience from interval accumulation: ENAF exhibits remarkable tolerance to temporal jitter and spatial false events because random spurious triggers lack angular correlation and average out when polarities are accumulated over interval maps.
Superior training efficiency and convergence: ENAF achieves 30 dB PSNR within 5 minutes using 5 views, while SAX-NeRF requires over 60 minutes and 25 views to achieve comparable quality, demonstrating massive data and compute efficiency.

Highlights & Insights¶

Physics-grounded multimodal integration: Maps the logarithmic intensity differences of event cameras directly onto line-integral attenuation differences derived from Beer–Lambert's law, forming an exact physical bridge between neuromorphic vision and 3D tomography.
Calibration-free spatial pattern matching: Bypasses the need for precise physical threshold calibration through normalized pattern loss, making the framework immediately practical on noisy experimental hardware.
Broad cross-task transferability: The principle of anchoring absolute scale with sparse frames while guiding continuous derivatives with asynchronous events readily extends to 4D cardiac angiography, neutron tomography, and ultra-high-speed industrial inspection.

Limitations & Future Work¶

Limited real-world sample diversity: Real-world evaluation currently centers on a single nano-architected lattice specimen scanned at a synchrotron beamline, leaving complex soft-tissue anatomical organs and heterogeneous dense materials to future validation.
Static rigid assumption: The forward model assumes rigid, uniform rotation without explicit motion compensation for non-rigid anatomical deformations like respiration or cardiac pulsation.
Future directions: Integrating temporal deformation fields to achieve 4D dynamic CT from microsecond event streams, paired with rigorous clinical dosimetry assessments.

vs NAF / SAX-NeRF (Frame-only INR CT): Frame-based INRs struggle severely with streak artifacts when limited to 5–10 views due to under-constrained ray sampling; ENAF leverages asynchronous event streams to supply continuous angular supervision between sparse views.
vs EventNeRF / E-NeRF (Optical Event Radiance Fields): Optical event NeRF models operate on surface radiance emission and reflection in 3D scenes; ENAF establishes the first formulation for volumetric line-integral attenuation in X-ray penetrative physics.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ Establishes the first neuromorphic X-ray computed tomography framework and proves the concept on real synchrotron beamline hardware.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Encompasses four synthetic anatomical datasets, extensive noise and temporal jitter ablations, and real-world beamline experiments.
Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally rigorous mathematical and physical derivations, well-structured figures, and clear empirical analysis.
Value: ⭐⭐⭐⭐⭐ Opens a highly promising hardware-software codesign pathway for ultra-low-dose medical CT and ultra-fast dynamic non-destructive testing.