Research map / Learning theory and algorithms / Prediction and decision losses
ActSort: An active-learning accelerated cell sorting algorithm for large-scale calcium imaging datasets
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.
Introduces ActSort, an active-learning cell sorting pipeline with a memory-efficient GUI, 76 engineered features, and a new DCAL query algorithm, validated on a ~160,000-annotation calcium imaging benchmark.
Problem
Manual curation (cell sorting) of false positives from automated cell extraction in large-scale calcium imaging datasets is laborious and subject to annotator inconsistency and bias.
Approach
Combines expert-engineered spatiotemporal features, a memory-compressed data format, and a new discriminative-confidence active learning query (DCAL) that adaptively mixes confidence-based and discriminative queries to select outlier boundary cells for annotation with linear classifiers.
Key results
- DCAL reached human-level balanced accuracy with about 3% of cell candidates annotated when training from scratch, versus about 10% for random sampling.
- With pre-labeled samples, DCAL reached human-level balanced accuracy with as little as 1% annotated and converged at under 6%, whereas random sampling needed about 8%.
- Feature engineering raised the false-positive rejection accuracy from 67% to 72% relative to traditional features; individual human annotator ICC was 0.56 versus 0.79 for majority-vote labels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add 76 expert-engineered features to the traditional feature set (ablation) | Linear cell classifiers · Hemisphere datasets, 3 mice | false-positive rejection accuracy: 67% -> 72% | Helped |
| Add 76 expert-engineered features to the traditional feature set (ablation) | Linear cell classifiers · Hemisphere datasets, 3 mice | true-positive acceptance accuracy: 97% -> 97% | No clear effect |
| Use DCAL instead of random sampling | ActSort · Hemisphere datasets, 3 mice | annotation fraction to reach human-level balanced accuracy: roughly 10% -> roughly 3% | Helped |
| Use DCAL instead of random sampling | ActSort on combined hemisphere datasets · Combined datasets across 3 mice | balanced accuracy at sparse annotation: < 1% | Helped |
| Use DCAL instead of CAL or DAL (ablation) | ActSort on combined hemisphere datasets · Combined datasets across 3 mice | balanced accuracy | Helped |
| Initialize DCAL with previously annotated samples | ActSort · New hemisphere dataset from a different mouse | annotation fraction to reach human-level balanced accuracy: c.a. 8% -> 1% | Helped |
| Use DCAL instead of random sampling with pre-labeled data (ablation) | ActSort · New hemisphere dataset from a different mouse | balanced accuracy: c.a. 8% -> 1% | Helped |
| Use ResNet-50 features instead of the expert-engineered feature set (ablation) | Linear cell classifier · Hemisphere datasets | true-positive rate: down to ∼90% | Hurt |
| Vary DCAL's initial trade-off weight (ablation) | ActSort DCAL · Cell-sorting benchmarks | balanced accuracy | No clear effect |
| Train the classifier rather than rely on an individual human annotator | Cell classifiers · Five-mouse annotation benchmark | annotation consistency (ICC): 0.56 ± 0.06 -> 0.64 ± 0.07 | Helped |
| Annotate randomly instead of selecting with DCAL (ablation) | ActSort · Hemisphere datasets, 3 mice | annotation fraction to converge: < 10% -> 50% | Hurt |
Surprising findings
- Individual human annotators were inconsistent (ICC 0.56), and human annotators were overconfident relative to annotator agreement, while trained classifiers were better calibrated.
- CAL and DAL alone performed poorly relative to random sampling when multiple animals were combined, while DCAL remained strong.
Limitations
- Depends on upstream cell extraction; extraction errors propagate into sorting.
- Uses linear classifiers only; deep models not explored.
- Validated only for binary sorting on one-photon and two-photon calcium imaging; annotator expertise effects not fully explored.
Experimental setup
Datasets: five calcium imaging movies (three one-photon hemisphere, one cortex-wide one-photon from prior work, one two-photon V1) striatum recordings from prior work (ICA-extracted)
Benchmarks: ActSort annotation benchmark (five movies, ~40,000 cells, ~160,000 annotations) published cortex-wide one-photon dataset striatum dSPN/iSPN dataset
Metrics: balanced accuracy true positive rate true negative rate area under ROC curve intraclass correlation
Baselines: random sampling CAL DAL traditional features classifier ResNet-50 features CAIMAN classifier
Builds on: confidence-based active learning discriminative active learning EXTRACT CAIMAN
Closest papers by meaning
- SimSort: A Data-Driven Framework for Spike Sorting by Large-Scale Electrophysiology Simulation (NeurIPS 2025 · similarity 0.86)
- realSEUDO for real-time calcium imaging analysis (NeurIPS 2024 · similarity 0.84)
- ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy (ICML 2025 · similarity 0.83)
- SpikeSSL: A Universal Spike Inference Framework with Dynamics-Informed State-Space Layers (NeurIPS 2026 · similarity 0.83)
- Retrospective for the Dynamic Sensorium Competition for predicting large-scale mouse primary visual cortex activity from videos (NeurIPS 2024 · similarity 0.82)
- Inpainting the Neural Picture: Inferring Unrecorded Brain Area Dynamics from Multi-Animal Datasets (NeurIPS 2025 · similarity 0.82)
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.