Research map / Time series / Foundation models for time series
A Compact Implicit Neural Representation for Efficient Storage of Massive 4D Functional Magnetic Resonance Imaging
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an INR-based per-subject compressor for 4D fMRI that models ICA-initialized reusable neuronal activation patterns and their spatial distributions with INR networks and a feature fusion block, outperforming H.264/H.265/JPEG/NeRF/SIREN/HNeRV/SSF/DVC at ~100x compression.
Problem
Compressing massive 4D fMRI data is hard due to complex temporal dynamics, low SNR, and redundancies within and between brain regions that frame-adjacent video codecs and INR video compressors do not exploit.
Approach
Model fMRI as a weighted superposition of K learnable reusable activation patterns (a learnable matrix initialized by ICA), with INR MLPs predicting each pattern's spatial distribution from coordinates, followed by a feature fusion network, rather than directly fitting coordinates-to-signal mappings.
Key results
- At ~100x compression (127.83x), ICAINR reaches 79.31 dB PSNR versus 70.49 dB for HNeRV and 67.63 dB for NeRF, the next best methods.
- FLA mean residual of 0.32 at ~128x versus 0.37 for NeRF (second best) and 0.86-1.32 for other codecs and INR baselines.
- Classification accuracy of 93.89% and AUC 0.9975 at compression below 100x, close to ground truth.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Feature Fusion Block (ablation) | ICAINR · fMRI ablation study | PSNR (dB): 78.67 -> 78.74 | Helped |
| initialize activation patterns with ICA instead of uniform initialization (ablation) | ICAINR · fMRI ablation study | FLA Residual: 0.36$\pm$0.30 -> 0.32$\pm$0.24 | Helped |
| initialize activation patterns with ICA instead of normal initialization (ablation) | ICAINR · fMRI ablation study | FLA Residual: 0.49$\pm$0.35 -> 0.32$\pm$0.24 | Helped |
| decompose fMRI signals into neuronal activation patterns rather than directly model the 4D data with INR (ablation) | ICAINR · fMRI ablation study | PSNR (dB): 72.93 -> 78.67 | Helped |
| use ICAINR rather than existing compressors at similar compression ratios | fMRI · fMRI (≈100× compression) | PSNR (dB): 70.49 -> 79.31 | Helped |
Surprising findings
- Traditional codecs H.264/H.265 and JPEG attained lower PSNR than per-frame INR methods such as NeRF and SIREN at similar compression, and the SSF data-driven codec nearly fails on fMRI (33.03 dB PSNR).
Limitations
- Per-subject network training is slow compared with conventional compressors.
- All ICA components are used, including ones with little neural information, so the model is not compact.
- Model compression limited to standard quantization and Huffman coding without tailored techniques.
Experimental setup
Datasets: OpenfMRI ds000007 OpenfMRI ds000101 OpenfMRI ds000102 Haxby
Benchmarks: OpenfMRI ds000007 OpenfMRI ds000101 OpenfMRI ds000102 Haxby
Metrics: PSNR SSIM compression ratio FLA residual mean/std FCA residual mean/std 10-fold classification accuracy AUC
Baselines: JPEG H.264 H.265 SSF DVC NeRF HNeRV SIREN
Builds on: NeRF SIREN HNeRV
Closest papers by meaning
- SCI: A Spectrum Concentrated Implicit Neural Compression for Biomedical Data (AAAI 2023 · similarity 0.88)
- Grids Often Outperform Implicit Neural Representation at Compressing Dense Signals (NeurIPS 2025 · similarity 0.86)
- SINR: Sparsity Driven Compressed Implicit Neural Representations (CVPR 2025 · similarity 0.86)
- FAST-Brain: A Flow-Aligned Spatio-Temporal Surrogate Brain Model (NeurIPS 2026 · similarity 0.86)
- Mind Reader: Reconstructing complex images from brain activities (NeurIPS 2022 · similarity 0.85)
- A sampling theory perspective on activations for implicit neural representations (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.