Research map / Probabilistic and scientific modeling / Dynamical systems
realSEUDO for real-time calcium imaging analysis
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.
Builds a real-time, frame-by-frame calcium imaging demixing method (realSEUDO) by speeding up the SEUDO robust time-trace estimator and adding an online cell-finding loop, and reports higher speed and more detected cells than OnACID.
Problem
Real-time, frame-by-frame extraction of single-neuron fluorescence time traces from multi-photon calcium imaging for closed-loop experiments, without initialization data and robust to unidentified neurons.
Approach
Speed up the SEUDO sparse-emulation robust estimator (C++ FISTA with per-dimension momentum reset, approximate Lipschitz constant, fewer Gaussian kernels, patch-based parallelism) and add a feedback loop that promotes residual blobs from temporary to stable profiles, so cells are discovered online starting from an empty set.
Key results
- On NAOMi simulation, realSEUDO found 201 true cells vs. 152 for OnACID and 308 for CNMF, processing 67.8 fps end-to-end vs. 9.2 fps for OnACID.
- On 36 mouse CA1 recordings, realSEUDO averaged 162 fps vs. 26 fps for OnACID and 13 fps for CNMF (6.5x and 12.5x speedups).
- Sparse SEUDO kernel reduction gave a 29-34.5x speedup over original SEUDO per cell, reducing runtime from ~5.8-6.9 s/cell to 0.2 s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use realSEUDO instead of OnACID | NAOMi simulated video cell detection · NAOMi simulated video | true cells found: 152 -> 201 | Helped |
| Use realSEUDO instead of OnACID | NAOMi simulated video processing · NAOMi simulated video | processing speed (frames per second): 9.2 -> 67.8 | Helped |
| Use realSEUDO instead of OnACID | Mouse CA1 recordings · Mouse CA1 recordings (36 movies) | processing speed (frames per second): 26 -> 162 | Helped |
| Use realSEUDO instead of CNMF | Mouse CA1 recordings · Mouse CA1 recordings (36 movies) | processing speed (frames per second): 13 -> 162 | Helped |
| Replace MATLAB/TFOCS implementation with optimized C++ implementation | SEUDO · SEUDO benchmark (45000 frames, 50 cells) | runtime per cell: ~5.8-6.9 s/cell -> 0.9-1.1 s/cell | Helped |
| Use sparse Gaussian kernels | SEUDO · SEUDO benchmark (45000 frames, 50 cells) | runtime per cell: 0.9-1.1 s/cell -> 0.2 s | Helped |
| Add momentum stopping when the gradient changes sign (ablation) | FISTA on handwritten digit recognition · Handwritten digit recognition | error: 0.1662 -> 0.0699 | Helped |
| Fix η=1 in addition to gradient-sign momentum stopping (ablation) | FISTA on handwritten digit recognition · Handwritten digit recognition | error: 0.0699 -> 0.0701 | No clear effect |
| Use a 4x descent rate with η=1 and gradient-sign momentum stopping (ablation) | FISTA on handwritten digit recognition · Handwritten digit recognition | error: 0.0701 -> 0.0751 | Hurt |
| Automatically adjust the rate with gradient-sign momentum stopping (ablation) | FISTA on handwritten digit recognition · Handwritten digit recognition | error: 0.0952 -> 0.0631 | Helped |
| Set η=1 with auto-adjusted rate and gradient-sign momentum stopping (ablation) | FISTA on handwritten digit recognition · Handwritten digit recognition | error: 0.0631 -> 0.0654 | Hurt |
Surprising findings
- Reducing the number of Gaussian kernels (one per pixel to sparser spacing) gave over 100x speedup for kernels of diameter 30 pixels without substantial loss of false-transient removal.
- On the handwritten digit neural network test, the momentum-stop FISTA variant reached the same error as gradient descent in about 10x fewer passes, suggesting generality beyond neuroscience (though stochastic methods still performed faster).
Limitations
- Realtime benchmarks sensitive to compute environment (CPU frequency governor must be set to performance mode).
- Only cell detection is evaluated; motion correction is assumed from OnACID and the full pipeline is not in C++.
- Lower true-positive rate (75%) and 24% false-positive rate on transient classification compared to prior SEUDO work that averaged 3 frames.
Experimental setup
Datasets: NAOMi simulated video (20000 frames, 500x500) Gauthier et al. 2022 mouse CA1 in-vivo recordings (36 videos, 90x90, 41750 frames) OnACID demo mesoscope video (2000 frames) handwritten digits zip code dataset (MNIST-like 8x8, 7291 images) for FISTA test
Benchmarks: NAOMi simulation Gauthier et al. 2022 mouse CA1 recordings OnACID demo mesoscope video handwritten digits neural network training
Metrics: true positives false positives false negatives frames per second CPU seconds per frame temporal correlation with ground truth true/false transient retention rate
Baselines: OnACID CNMF FIOLA original SEUDO (MATLAB/TFOCS)
Builds on: SEUDO FISTA OnACID CNMF
Closest papers by meaning
- OnACID: Online Analysis of Calcium Imaging Data in Real Time (NeurIPS 2017 · similarity 0.88)
- Active learning of neural population dynamics using two-photon holographic optogenetics (NeurIPS 2024 · similarity 0.85)
- ActSort: An active-learning accelerated cell sorting algorithm for large-scale calcium imaging datasets (NeurIPS 2024 · similarity 0.84)
- SpikeSSL: A Universal Spike Inference Framework with Dynamics-Informed State-Space Layers (NeurIPS 2026 · similarity 0.83)
- convSeq: Fast and Scalable Method for Detecting Patterns in Spike Data (ICML 2024 · similarity 0.81)
- SimSort: A Data-Driven Framework for Spike Sorting by Large-Scale Electrophysiology Simulation (NeurIPS 2025 · similarity 0.81)
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.