Research map / Detection and segmentation / Open-set and few-shot detection
FISBe: A Real-World Benchmark Dataset for Instance Segmentation of Long-Range Thin Filamentous Structures
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.
Releases FISBe, a 101-volume expert-annotated 3D light microscopy dataset of Drosophila neurons for instance segmentation of thin filamentous structures, with new centerline-based metrics and three baselines.
Problem
Instance segmentation of long-range, thin, filamentous and intertwined neurons in volumetric light microscopy lacks real-world benchmarks; existing evaluations rely on synthetic data and standard IoU-based metrics are ill-suited to thin structures.
Approach
A real-world 3D multi-color light microscopy dataset with pixel-wise neuron instance masks, plus clDice-based instance F1, a centerline-recall ground truth coverage metric, and an aggregate score S = 0.5 avF1 + 0.5 C, with false split/merge counts.
Key results
- PatchPerPix achieved the best overall results on the full test set, with S 0.35 and avF1 0.34 (per the Table 1 values); all baselines reached maximum avF1 of about 0.34.
- Duan et al. color clustering had the most false splits (45 on test) while FFN+partly reached S 0.27 on test.
- Dataset comprises 101 3D volumes (30 completely labeled, 71 partly labeled) with 590 annotated neurons.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add partly labeled training data (ablation) | FFN · FISBe validation | S: 0.25±0.01 -> 0.27±0.01 | Helped |
| add partly labeled training data (ablation) | FFN · FISBe test | S: 0.25±0.03 -> 0.27±0.01 | Helped |
Surprising findings
- The non-learnt color clustering baseline of Duan et al. achieved test avF1 comparable to learnt methods, and it had the lowest false merge count.
- Pixel-level IoU-based mAP was judged unsuitable for thin structures, motivating clDice-based metrics.
Limitations
- Bias toward sparser expression density samples, so denser samples are not represented.
- Baselines handle no or only small overlaps and are computationally demanding.
- Partly labeled images make F1 scores of completely and partly labeled subsets not directly comparable.
Experimental setup
Datasets: FlyLight MCFO 40x Gen1 images
Benchmarks: FISBe
Metrics: avF1 ground truth coverage C aggregate score S clDice_TP false splits false merges tp
Baselines: PatchPerPix Flood Filling Networks Duan et al. color clustering
Builds on: Metrics Reloaded clDice PatchPerPix Flood Filling Networks
Closest papers by meaning
- Efficient Connectivity-Preserving Instance Segmentation with Supervoxel-Based Loss Function (AAAI 2025 · similarity 0.88)
- Learning Multimodal Volumetric Features for Large-Scale Neuron Tracing (AAAI 2024 · similarity 0.85)
- Cross-Dimension Affinity Distillation for 3D EM Neuron Segmentation (CVPR 2024 · similarity 0.84)
- Micro-Bench: A Microscopy Benchmark for Vision-Language Understanding (NeurIPS 2024 · similarity 0.84)
- From Pixels to Views: Learning Angular-Aware and Physics-Consistent Representations for Light Field Microscopy (NeurIPS 2025 · similarity 0.83)
- MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology Assays (NeurIPS 2026 · similarity 0.83)
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.