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

Agnieszka Pregowska

cs.CVMar 26, 2026

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

Authors: Agnieszka Pregowska

Organizations: Institute of Fundamental Technological Research Polish Academy of Sciences, Pawinskiego 5B, Warsaw, 02-106, Poland

Abstract

Implicit neural representations (INRs) offer continuous coordinate-based encodings for atlas registration, cross-modality resampling, sparse-view completion, and compact sharing of neuroanatomical data. Yet reproducible evaluation is lacking for high-resolution larval zebrafish microscopy, where preserving neuropil boundaries and fine neuronal processes is critical. We present a reproducible INR benchmark for the MapZebrain larval zebrafish brain atlas. Using a unified, seed-controlled protocol, we compare SIREN, Fourier features, Haar positional encoding, and a multi-resolution grid on 950 grayscale microscopy images, including atlas slices and single-neuron projections. Images are normalized with per-image (1,99) percentiles estimated from 10% of pixels in non-held-out columns, and spatial generalization is tested with a deterministic 40% column-wise hold-out along the X-axis. Haar and Fourier achieve the strongest macro-averaged reconstruction fidelity on held-out columns (about 26 dB), while the grid is moderately behind. SIREN performs worse in macro averages but remains competitive on area-weighted micro averages in the all-in-one regime. SSIM and edge-focused error further show that Haar and Fourier preserve boundaries more accurately. These results indicate that explicit spectral and multiscale encodings better capture high-frequency neuroanatomical detail than smoother-bias alternatives. For MapZebrain workflows, Haar and Fourier are best suited to boundary-sensitive tasks such as atlas registration, label transfer, and morphology-preserving sharing, while SIREN remains a lightweight baseline for background modelling or denoising.

Figures & tables

Model	Macro-Mean	Macro-Std	Micro-Mean	Wins
Haar	26.09	27.70	12.83	354
Fourier	25.95	26.59	13.75	288
SIREN	13.90	2.28	13.41	233
Grid	24.12	28.31	10.98	75

Table 1: Macro- and micro-averaged

\mathrm{PSNR}_{\text{test\_mean}}

(dB) across 950 images.

Regime	Haar	Fourier	SIREN	Grid
all-in-one	11.23	12.47	12.71	9.80
regions	26.47	26.57	17.68	26.45

Table 2: Area-weighted (micro)

\mathrm{PSNR}_{\text{test}}

(dB) by regime.

Figure 1: Overall distribution of $\mathrm{PSNR}_{\text{test\_mean}}$ for the four INR variants (all images). Fourier and Haar have similar medians but a broader upper tail. SIREN is the most concentrated.

Figure 2: Mean $\mathrm{PSNR}_{\text{test\_mean}}$ (macro) across methods. Haar and Fourier achieve the highest averages.

Figure 3: Box plot of $\mathrm{PSNR}_{\text{test\_mean}}$ showing broad distributions with occasional outliers.

Figure 4: Violin plot of $\mathrm{PSNR}_{\text{test\_mean}}$ illustrating per-image variability.

Figure 5: Representative reconstructions for selected MapZebrain regions using different positional encodings. Fourier and Haar preserve morphological detail and boundaries, SIREN smooths fine structures, and Grid exhibits strong blurring and vertical artifacts. Rightmost panels show column-wise MAE profiles.

Regime	Model	Macro-Mean	Macro-Std	Wins
all-in-one	Haar	12.41	2.53	291
all-in-one	Fourier	12.93	1.79	198
all-in-one	SIREN	12.70	0.82	228
all-in-one	Grid	10.25	2.15	6
regions	Haar	69.68	26.36	63
regions	Fourier	67.43	26.22	90

Table 3: Per-regime results: macro-averaged

\mathrm{PSNR}_{\text{test\_mean}}

(dB) and wins per method.

Model	SSIM ${}_{\text{test}}$ (↑)	Edge MAE ${}_{\text{test}}$ (↓)	PSNR ${}_{\text{micro}}$ [dB]
Fourier	0.689	0.183	13.80
Grid	0.687	0.246	10.93
Haar	0.634	0.213	12.83
SIREN	0.443	0.208	13.43

Table 4: Complementary test-only metrics on held-out columns (all images). Higher is better for SSIM; lower is better for Edge MAE. PSNR

{}_{\text{micro}}

reported for reference. Macro averages across images.

Figure 6: Smooth bias characteristic of sinusoidal INRs (SIREN).

Figure 7: Example reconstruction under column-wise holdout regime illustrating directional overfitting.

Figure 8: Haar positional encoding preserves sharp edges but introduces block boundary artifacts.

Setting	Value
Intensity normalization	per-image percentiles $(1,99)$
Hold-out split	blocked_cols_X , $\alpha=0.40$
Train subsampling	$10\%$ pixels from training columns
Epochs / Optimizer	$10$ / AdamW ( $\mathrm{lr}=10^{-3}$ , $\mathrm{wd}=10^{-6}$ )
Loss	Smooth- $\ell_{1}$ ( $\beta=0.01$ )
Batch size	$\min(131{,}072,\;\max(65{,}536,\;\lfloor HW\cdot 0.10/2\rfloor))$

Table 5: Key hyperparameters

Model	Mean $\pm$ SD	Median	IQR	N
Fourier	25.95 ± 26.59	13.16	[11.87, 18.05]	950
Grid	24.12 ± 28.31	9.99	[9.36, 16.43]	950
Haar	26.09 ± 27.70	14.10	[10.31, 16.92]	950
SIREN	13.90 ± 2.28	13.00	[12.21, 15.18]	950

Table 6: Per-method descriptive statistics of

\mathrm{PSNR}_{\text{test\_mean}}

(dB) across all images (N=950): mean

\pm

SD, median, and IQR.

Regime	Model	Mean $\pm$ SD	Median	IQR	N
all-in-one	Fourier	12.93 ± 1.79	12.47	[11.73, 13.69]	723
all-in-one	Grid	10.25 ± 2.15	9.80	[9.06, 10.32]	723
all-in-one	Haar	12.41 ± 2.53	12.96	[10.18, 14.52]	723
all-in-one	SIREN	12.70 ± 0.82	12.54	[12.08, 13.23]	723
regions	Fourier	67.43 ± 26.22	55.81	[45.77, 100.00]	227
regions	Grid	68.29 ± 27.84	54.89	[44.36, 100.00]	227

Table 7: Per-method descriptive statistics by regime (macro view): mean

\pm

SD, median, and IQR.

Comparison	$p$ -value (Wilcoxon)	Cliff’s $\delta$	95% CI	Paired N
Haar vs. SIREN	2.617e-14	0.051	[-0.008, 0.103]	950
Fourier vs. SIREN	2.940e-34	0.029	[-0.020, 0.079]	950
Haar vs. Grid	1.692e-49	0.308	[0.254, 0.359]	950
Fourier vs. Grid	5.293e-67	0.435	[0.387, 0.487]	950
Haar vs. Fourier	1.432e-08	-0.053	[-0.107, 0.001]	950

Table 8: Pairwise comparisons: macro PSNR.

Model	Outliers	Total (regions)	Share (%)
Fourier	73	227	32.2
Haar	67	227	29.5
Grid	94	227	41.4
SIREN	0	227	0.0

Table 9: Outliers (PSNR

\geq 100

dB) in regions .

Regime	Best method	Wins	N	Share (%)
all-in-one	Fourier	198	723	27.4
all-in-one	Grid	6	723	0.8
all-in-one	Haar	291	723	40.2
all-in-one	SIREN	228	723	31.5
regions	Fourier	90	227	39.6
regions	Grid	69	227	22.5

Table 10: Wins per regime with shares.

Figure 9: Number of per-image wins in the all-in-one regime. Haar dominates, SIREN remains competitive, the grid almost never wins.

Figure 10: Number of per-image wins in the regions regime. Explicit encodings (Fourier, Haar, Grid) dominate over SIREN.

Model	Median	IQR	Trimmed Mean (10%)
Haar	14.10	[10.31, 16.92]	16.84
Fourier	13.16	[11.87, 18.05]	16.27
Grid	9.99	[ 9.36, 16.43]	14.22
SIREN	13.00	[12.21, 15.18]	13.62

Table 11: Robust PSNR statistics across all images.

Explore similar work

May 11, 2026

INFANiTE: Implicit Neural representation for high-resolution Fetal brain spatio-temporal Atlas learNing from clinical Thick-slicE MRI

Spatio-temporal fetal brain atlases are important for characterizing normative neurodevelopment and identifying congenital anomalies. However, existing atlas construction pipelines necessitate days for slice-to-volume reconstruction (SVR) to generate high-resolution 3D brain volumes and several additional days for iterative volume registration, thereby rendering atlas construction from large-scale cohorts prohibitively impractical. We address these limitations with INFANiTE, an Implicit Neural Representation (INR) framework for high-resolution Fetal brain spatio-temporal Atlas learNing from clinical Thick-slicE MRI scans, bypassing both the costly SVR and the iterative non-rigid registration steps entirely, thereby substantially accelerating atlas construction. Extensive experiments demonstrate that INFANiTE outperforms existing baselines in subject consistency, reference fidelity, intrinsic quality and biological plausibility, even under challenging sparse-data settings. Additionally, INFANiTE reduces the end-to-end processing time (i.e., from raw scans to the final atlas) from days to hours compared to the traditional 3D volume-based pipeline (e.g., SyGN), facilitating large-scale population-level fetal brain analysis. Code: https://github.com/hu2274898/INFANiTE

Xiaotian Hu, Mingxuan Liu, Hongjia Yang +10

Tsinghua University · University of Oxford · Sichuan University

Implicit Neural Representations Longitudinal Medical Image Analysis Neuroimaging Brain MRI

Ask PDF

Jul 23, 2026

Recurrent Sinusoidal INRs for Efficient High-Fidelity Representation

We study sinusoidal recurrence as an iterative mechanism for harmonic spectral enrichment in implicit neural representations (INRs). Our analysis reveals that sinusoidal activations induce a harmonic line spectrum, providing a spectral account of how recurrent unrolling enriches the effective spectral support. We realize this principle with a shared sinusoidal block that iteratively refines the latent representation. We empirically validate the resulting spectral behavior against feed-forward INRs, non-sinusoidal recurrent variants, and equilibrium-style sinusoidal models. Complementing this analysis, we evaluate the proposed architecture across image and 3D representation tasks. On RGB image benchmarks, our method achieves higher fidelity than feed-forward baselines with fewer parameters and fewer optimization steps, and it further transfers favorably to super-resolution, NeRF, and SDF tasks.

Hyunmin Cho, Jaejun Yoo, Kyong Hwan Jin

Department of Electrical Engineering, Korea University, Seoul, South Korea · Graduate School of Artificial Intelligence, UNIST, Ulsan, South Korea

Implicit Neural Representations

Ask PDF

Aug 31, 2026

CORAL: A Benchmark for Structure-aware and Brain-wide Neuron Reconstruction in Light Microscopy

Automatic neuron reconstruction from light microscopy images is a central problem in computational neuroanatomy. While recent methods have achieved encouraging results on local image blocks, it remains unclear whether such progress translates to reconstruction that is both structurally accurate and scalable to the whole-brain scale. We present CORAL, the first benchmark for structure-aware evaluation of automatic neuron reconstruction from light microscopy images at both local and whole-brain scales. Built on a high-quality whole-brain fMOST dataset with carefully curated annotations, CORAL establishes two progressive tasks: block-level reconstruction, which evaluates reconstruction methods under limited spatial context, and brain-wide reconstruction, which assesses complete neuron reconstruction at the whole-brain scale. To account for topological correctness beyond geometric distance similarity, we introduce a structure-aware metric based on fiber prediction. To further achieve complete neuron reconstruction across the entire brain, we develop a brain-wide neuron tracing framework that extends arbitrary local reconstruction methods to the whole-brain scale through an iterative local-to-global process. Using this benchmark, we provide the first structure-aware comparison of mainstream methods for local neuron reconstruction and further evaluate their performance in brain-wide reconstruction. Our results underscore the importance of structure-aware evaluation and the need for more robust methods for complete neuron reconstruction.

Zekang Yang, Jiamin Li, Zhenghua Li +3

Department of Computer Science and Technology, Institute for AI, BNRist, Tsinghua University, Beijing 100084, China · Tsinghua Laboratory of Brain and Intelligence (THBI), IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing 100084, China · School of Basic Medical Sciences, Tsinghua University, Beijing, 100084, China +2

Neuroimaging Microscopy Image Analysis

Ask PDF