Neuromorphic X-ray Computed Tomography

ECCV 2026

Hongjian Wang, Goran Lovric, Benjamín Béjar

ETH Zurich Swiss Light Source Swiss Data Science Center

Neuromorphic X-ray CT acquisition with sparse frames and asynchronous events — NeuroXCT combines sparse frame projections with asynchronous event streams. The sparse projections provide absolute structural information, while events capture high-temporal- resolution angular changes between frames.

Abstract

X-ray computed tomography reconstruction is inherently an ill-posed inverse problem, particularly under sparse-view projections. Unlike traditional frame-based sensors, neuromorphic event cameras operate asynchronously at the pixel level and trigger events only when log-intensity changes exceed a threshold. This asynchronous sensing naturally captures high-frequency angular variations during object rotation, providing complementary information between sparsely sampled projections. We introduce a Neuromorphic X-ray CT framework and propose an Event-enhanced Neural Attenuation Field (ENAF) that jointly leverages sparse projections and event streams. Experiments on synthetic and real-world datasets show that ENAF improves sparse-view reconstruction quality over frame-only methods while maintaining practical training efficiency.

The Proposed Method

ENAF represents the volumetric attenuation field as a continuous neural function. During training, projection supervision anchors the absolute attenuation scale, while event supervision compares predicted projection differences against accumulated event maps over arbitrary angular windows.

Overview of the Event-enhanced Neural Attenuation Field method

Results

ENAF suppresses streak artifacts and preserves fine structures in both synthetic sparse-view CT benchmarks and a real-world nano-architected lattice dataset acquired at the TOMCAT beamline.

Qualitative comparison of 5-view sparse-view CT reconstruction — 5-view sparse-view CT reconstruction on synthetic datasets.

Qualitative comparison of 25-view sparse-view CT reconstruction — 25-view sparse-view CT reconstruction on synthetic datasets.

Qualitative comparison on the real-world lattice dataset — Real-world lattice reconstruction using 5, 10, and 25 frame projections.

Reference 3D reconstruction of the real-world lattice dataset.

Central slice of the reference real-world lattice reconstruction.

Event-only and event-enhanced reconstruction on the real-world lattice dataset — Event-only and event-enhanced reconstructions on the real-world lattice dataset.

Reconstruction quality versus training time.

BibTeX

@inproceedings{wang2026neuroxct,
  title     = {Neuromorphic X-ray Computed Tomography},
  author    = {Wang, Hongjian and Lovric, Goran and Béjar, Benjamín},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}