Implicit Neural Representations for Larval Zebrafish Brain Microscopy: A Reproducible Benchmark on the MapZebrain Atlas
The recent advancement in implicit neural representations (INRs) has opened new avenues for neuroanatomical data processing and analysis. A study published on arXiv (arXiv:2603.26811v1) presents a reproducible benchmark for the MapZebrain larval zebrafish brain atlas, focusing on high-resolution microscopy images. This work aims to enhance the understanding of neuronal processes by providing continuous coordinate-based encodings, which are crucial for various applications such as atlas registration, cross-modality resampling, and sparse-view completion.
Importance of Reproducible Evaluation
Despite the potential of INRs, the need for reproducible evaluations in high-resolution larval zebrafish microscopy has been a significant challenge. Preserving neuropil boundaries and fine neuronal structures is critical for accurate neuroanatomical analysis. The presented benchmark addresses this gap by evaluating different encoding methods on a standardized dataset.
Methodology
The benchmark utilizes a unified, seed-controlled protocol to compare various encoding methods, including:
- SIREN (Sinusoidal Representation Networks)
- Fourier features
- Haar positional encoding
- Multi-resolution grid
These methods were assessed on a dataset comprising 950 grayscale microscopy images, which included both atlas slices and single-neuron projections. The images were normalized using per-image (1,99) percentiles derived from 10% of the pixels in non-held-out columns. To test spatial generalization, a deterministic 40% column-wise hold-out along the X-axis was applied.
Results and Discussion
The findings revealed that Haar and Fourier encodings achieved the highest macro-averaged reconstruction fidelity on the held-out columns, with approximately 26 dB. The multi-resolution grid performed moderately, while SIREN, despite lower macro averages, remained competitive in area-weighted micro averages within the all-in-one regime.
Furthermore, the study employed Structural Similarity Index (SSIM) and edge-focused error metrics, indicating that Haar and Fourier methods preserved boundaries with greater accuracy compared to their smoother-bias alternatives. This suggests that explicit spectral and multiscale encodings are superior in capturing high-frequency neuroanatomical details.
Conclusion
The results of this benchmark highlight the potential of Haar and Fourier encodings for boundary-sensitive tasks such as atlas registration, label transfer, and morphology-preserving sharing within the MapZebrain workflows. On the other hand, SIREN can serve as a lightweight baseline for applications involving background modeling or denoising.
In conclusion, this study not only emphasizes the significance of reproducible benchmarks in neuroanatomical data analysis but also provides valuable insights into the strengths and weaknesses of various implicit neural representation techniques in the context of larval zebrafish brain microscopy.