Implicit neural representations for larval zebrafish brain microscopy: a reproducible benchmark on the MapZebrain atlas

Researcher Agnieszka Pregowska has published the first reproducible benchmark for Implicit Neural Representations (INRs) applied to high-resolution larval zebrafish brain microscopy. INRs are AI models that use neural networks to create continuous, coordinate-based encodings of complex data, which is crucial for tasks like registering brain atlases, resampling data across different imaging modalities, completing sparse views, and sharing neuroanatomical data in a compact form. The study addresses a critical gap in neuroscience AI, where preserving the fine boundaries of neuropil and delicate neuronal processes is essential for accurate analysis.

The benchmark employed a rigorous, seed-controlled protocol to compare four leading INR architectures: SIREN, Fourier features, Haar positional encoding, and a multi-resolution grid. It evaluated them on a dataset of 950 grayscale images from the MapZebrain atlas, including full atlas slices and single-neuron projections. A key test was spatial generalization, using a deterministic 40% column-wise hold-out along the X-axis to see how well models reconstructed unseen data.

The results were decisive. Models with explicit spectral and multi-scale encodings—Haar and Fourier features—outperformed others, achieving the strongest macro-averaged reconstruction fidelity on held-out columns at approximately 26 dB. They also excelled in preserving critical neuroanatomical boundaries, as measured by SSIM and edge-focused error metrics. In contrast, the popular SIREN model performed worse in macro averages but remained competitive in area-weighted micro averages, suggesting it is better suited as a lightweight baseline for tasks like background modeling.

This work provides a vital roadmap for computational neuroscience. It clearly indicates that for boundary-sensitive MapZebrain workflows—such as atlas registration, label transfer, and morphology-preserving data sharing—researchers should prioritize Haar or Fourier feature INRs. The reproducible benchmark sets a new standard for evaluating AI in neuroimaging, moving the field beyond anecdotal evidence toward controlled, comparative analysis.