Biology & medicine

Fruit Fly Neuron Reconstruction

Reconstruct individual neurons from 3D electron microscopy.

Task & evaluation

Objective: Reconstruct individual neurons while preserving their long-range continuity.
Input: 3D MaleCNS electron microscopy on a 16 nm voxel grid.
Output: An instance volume: one neuron ID per voxel; zero denotes unassigned voxels.
Data: Public MaleCNS images and reference-derived supervision; held-out spatial regions for evaluation.
Evaluation: Normalized expected run length (nERL) along reference skeletons; higher is better.
Research challenge: Preserve thin branches and cell identity across large volumes without false merges.

Understanding the task

Follow the data through reconstruction, evaluation and the measured baseline experiment.

01Data and reconstruction task

A voxel is a pixel in 3D. Each predicted instance ID groups voxels belonging to the same neuron.

MODEL INPUT

3D EM image

A volume of grayscale electron-microscopy measurements.

→

MODEL OUTPUT

Instance volume

One neuron ID per voxel. Zero marks omitted or unassigned voxels.

→

EVALUATION

Given reference skeleton

The scorer reads predicted IDs along fixed neuron centerlines derived from released reference segmentation.

A real view into the tissue.

MaleCNS electron microscopy · 16 nm voxels · 8.192 μm field of view

Slice depth

Real MaleCNS electron microscopy slice through the center of the scoring region — Electron microscopy

Aligned FFN instance segmentation overlaid on the same electron microscopy slice — FFN prediction overlay

Colors indicate predicted neuron IDs; unassigned voxels retain their grayscale appearance. Both views show the same location. A 2D slice shows local boundaries; ERL evaluates continuity through the 3D volume.

Data: MaleCNS / FlyEM collaboration · CC BY 4.0 · Prediction: track FFN baseline.

Two inputs to the scorerPredicted instance volume + given reference skeleton. The reconstruction model receives the EM image; reference neuron identities belong to evaluation.

Where does the reference data come from?

MaleCNS means the male fruit fly’s central nervous system. The released EM and reconstruction dataset [4] comes from a collaboration involving FlyEM at HHMI Janelia, the University of Cambridge, MRC Laboratory of Molecular Biology and Google Research. Training uses reference-derived supervision; evaluation uses held-out spatial regions. Region-level separation does not guarantee disjoint neuron identities.

MaleCNS dataset

02How continuity is evaluated

Expected Run Length (ERL) measures the length of correctly reconstructed reference segments, weighted toward longer segments.

Predicted ID APredicted ID BPredicted ID C

The ideal denominator

Assign each reference point its true neuron ID and run the same scorer. For paths of length 6 and 4:

ERL_ideal = (6² + 4²) / (6 + 4) = 5.2

This is a reference-based normalization, computed directly from the given skeleton.

The intuition

Imagine a uniformly chosen point along the total reference path. A length-6 neuron receives 60% of the weight; a length-4 neuron receives 40%.

nERL = ERL_prediction / ERL_ideal

ERL is calculated from all reference lengths; the random-point view explains the weighting.

How splits, merges and omissions affect ERL

A predicted ID spanning two reference neurons is treated as a merge; edges assigned to that ID contribute zero correct length. Edges touching ID zero are omitted. Edges crossing predicted IDs are split. Correct edge lengths are grouped by reference-neuron/predicted-ID pair.

Writing Lᵢ for each reference neuron’s length and Cᵢ,ₛ for correct length in a group: ERL = Σ Cᵢ,ₛ² / Σ Lᵢ. [5]

03The reference pipeline

The pipeline combines an initial Flood-Filling Network (FFN) with learned proposals and shape-based decisions inspired by PATHFINDER. [1] [2]

FFN

Grow neuron fragments

Starting from seeds, the network repeatedly expands a neuron mask using image evidence and its current mask.

SENSE

Find candidate connections

A 3D segmentation network identifies potential continuation regions and proposes fragment pairs.

SHAPE

Score the combined geometry

A point-cloud classifier evaluates whether two fragments form a plausible neuron. Accepted pairs are merged. [3]

Scientific contributionThe track makes topology-sensitive neuron reconstruction a measurable learning-and-compute problem. PATHFINDER’s research contribution is using learned connection proposals and morphology to repair fragmented reconstructions. The track evaluates a simplified proposal-and-repair pipeline.

04Model architecture and scale

The models combine recurrent mask growth, dense 3D prediction and point-cloud classification. [6]

FFN

1.11M parameters

EM crop + current mask logits2 channels · 33 × 33 × 33

↓

3D convolution2 → 32 channels

↓

20 residual blocks2 × 3³ convolutions / block 32 channels · channel LayerNorm

↓

1 × 1 × 1 output convolutionAdd update to current mask logits

↻ Repeat as the mask grows through the volume

Small network, repeated spatial decisions. The 33³ network field of view moves inside each much larger FFN window.

SENSE

5.02M parameters

EM crop + segment prompt2 channels · 128 × 128 × 128

↓

128³ 8 channels

64³ 16

32³ 32

16³ 64

8³ 64

4³ bottleneck · 64 channels

↓ Decoder + skip connections ↑

128³ continuation logitsResidual 3D encoder–decoder

Large dense activations at every candidate center. Each crop contains about 58× as many voxels as one FFN crop.

SHAPE

8.19M parameters

8,192 × 3 coordinatesPhysical geometry · 10 μm radius

↓

Point embedding3 → 32 features

↓

6 set-abstraction stagesPoints: 8,192 → 128 Features: 32 → 2,048 Farthest-point sampling · 16 neighbors

↓

Global max pool → classifier2,048 → 512 → 256 → 1 logit

One plausibility score per candidate pair

SHAPE has the most parameters of the three models and evaluates a small candidate set. Sampling and nearest-neighbor search add geometric work.

05Spatial context and reconstruction windows

The scored center is 512³ voxels. Shape reasoning needs surrounding anatomy, so reconstruction covers 1762³ voxels on a 16 nm grid.

1762 × 1762 × 1762 voxelsThin outer blocks · 18 voxels along at least one axis

Rotate Separate blocks

125 follows from coverage. The 436-voxel core and 32-voxel margin are engineering settings. Neighboring instances are matched through overlap regions.

Why reconstruct 40.76× the scoring volume?

The SHAPE context radius is 10 μm, equivalent to 625 voxels on each side at 16 nm resolution. Thus 512 + 2 × 625 = 1762, and (1762 / 512)³ = 40.76. This anatomy context is distinct from the 32-voxel overlap used to stitch FFN windows.

06Simplifications relative to PATHFINDER

The track preserves the proposal-and-repair structure while reducing ensemble and combinatorial work.

Component	PATHFINDER paper pipeline	Track baseline
Initial reconstruction	3 FFN weight sets × 2 seed orders; consensus	1 corrected checkpoint; ascending seed order
Proposal comparison	Traditional FFN comparison uses independent proposals	FFN resegmentation and SHAPE receive the same 36 SENSE pairs
Merge search	Also evaluates combinations of 3 or 4 fragments	Disjoint binary merges; each initial fragment used at most once

BINARY MERGE EXAMPLE

A—B—C→ABC

Once A–B is accepted, B is occupied and B–C is skipped. Multi-fragment search can also evaluate the joint candidate {A, B, C}.

What does “ascending seed order” mean?

Seeds are positions where FFN begins growing an instance. Ascending order sorts coordinates lexicographically by (z, y, x) within each window. An earlier reconstruction can occupy voxels and change what later seeds can grow into. The descending variant reverses this list.

07Measured reconstruction results

One fixed test context and one corrected FFN checkpoint. Both repair branches use the same candidate list and reference scorer.

36candidate pairs

6SHAPE merges accepted

29.43%reference nodes omitted

Method	Normalized ERL ↑	Split count ↓	Merge count ↓
FFN baseline	0.0000367474	3,993	5,047
FFN + shared proposals	0.0000367474	3,993	5,047
FFN + SHAPE	0.0000366064	3,984	5,047

OBSERVED EFFECT

−0.3837% relative nERL

A small decrease after SHAPE. Fewer splits did not translate into higher length-weighted continuity. Stability across independent training seeds remains unmeasured.

LIKELY BOTTLENECK

Weak base + limited repair reach

Very low starting continuity and only 36 repair candidates limit the opportunity for recovery. Candidate distribution mismatch and false-positive merge decisions are hypotheses requiring per-candidate analysis.

STATE-OF-THE-ART CONTEXT

0.820 → 0.942

FFNv1.5 agglomeration → PATHFINDER, reported on the paper’s mouse IBEAM-mSEM benchmark. [1]

The paper demonstrates the potential of morphology-guided reconstruction repair. Its mouse benchmark provides research context; the fruit-fly evaluation uses different data and evaluation conditions.

08Interpreting the training signal

FFN / SENSE

Dice is foreground F1.

Dice = 2 |prediction ∩ truth| / (|prediction| + |truth|)

Selected validation Dice: FFN 0.8335 · SENSE 0.5231.

SENSE’s first six validations, through 24.25 min, had zero Dice at a 0.5 probability threshold. BCE still fell from 0.605 to 0.021; Dice reached 0.282 at 29.09 min. Probability learning can precede thresholded foreground overlap.

SHAPE

70.31% accuracy needs context.

Predict acceptPredict rejectValid pair320Invalid pair1913

45 of 64 balanced validation examples were correct. Only 13 of 32 invalid pairs were rejected.

One example changes accuracy by 1.56 percentage points. This diagnostic uses threshold 0.5; final merging uses 0.8. Accepted-merge precision at the deployment threshold needs separate measurement.

09Compute and memory

Training and downstream repair, starting from a precomputed FFN segmentation.

TRAINING · 4 A5000 GPUs

24.1 GPU-hours

About 6 hours elapsed across three sequential training stages.

FFNSENSESHAPE

FFN~3 h × 4 GPUs12.0 GPU-h

SENSE~2 h × 4 GPUs8.0 GPU-h

SHAPE~1 h × 4 GPUs4.0 GPU-h

Domain training: 24.0 GPU-h + 0.1 GPU-h overhead.

DOWNSTREAM REPAIR · A5000 GPUs

32.7 GPU-hours

1 h 26 min elapsed · up to 28 GPUs concurrently.

SENSE · 98.7%

SENSE7 jobs × 4 GPUs32.2 GPU-h

FFN resegmentation1 job × 4 GPUs0.2 GPU-h

SHAPE1 job × 4 GPUs0.2 GPU-h

GPU-hours = allocated GPUs × elapsed time.

INITIAL RECONSTRUCTION125 windows

FFN growth + overlap stitching

Precomputed segmentation

→

PROPOSAL SWEEP1.17M centers

SENSE evaluates eligible centers

32.2 GPU-h

→

PAIR EVALUATION36 pairs

Same list to both repair branches

0.4 GPU-h combined

Where the compute goes: a broad, dense 3D proposal sweep precedes a small pair-classification workload. Improving candidate scheduling, batching and feature reuse can target this measured bottleneck while retaining coverage.

Memory and storage

Resource	Usage	Configuration
FFN training GPU memory	4.8 GB peak allocated	Batch size 16 per GPU
SENSE training GPU memory	11.5 GB peak allocated	Batch size 8 per GPU
Storage footprint	71.5 GB sampled peak	Data, models, intermediate results and runtime files

10Research directions

Long-range identity

Train and validate for continuity across thin branches, ambiguous boundaries and window seams. Pair mask metrics with topology-aware validation.

Calibrated repair

Measure candidate recall and accepted-merge precision on actual predicted fragments. Study higher-order candidates and train on realistic failure cases.

Efficient inference

Batch FFN growth updates, schedule windows by estimated work, and reuse image features and candidate geometry. Verify reconstruction equivalence when optimizing.

Context and memory

Use chunked scoring, spatial indices and streamed point clouds. Smaller scoring cores alone save limited work when wide anatomy context dominates volume.

11References and provenance

Accelerating neuron reconstruction with PATHFINDER

Januszewski, Templier, Hayworth, Peale & Hess · bioRxiv, 2025

Source for SENSE, SHAPE, the multi-fragment reconstruction pipeline, and the paper’s mouse-benchmark comparison.

Paper / DOI Author institution

High-precision automated reconstruction of neurons with flood-filling networks

Januszewski et al. · Nature Methods, 2018

Foundation for iterative, mask-conditioned neuron growth. The track uses an independently reconstructed, 20-block FFN v1.5-style architecture.

Paper / DOI Google FFN License

PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies

Qian et al. · NeurIPS, 2022

Model-family reference for SHAPE. The six-stage classifier, widths and training setup are track-specific choices.

Paper PointNeXt OpenPoints MIT license

Sexual dimorphism in the complete Drosophila male central nervous system connectome

Berg et al. · Cell, 2026

Scientific context for the MaleCNS image and reconstruction dataset used to prepare this track’s supervision and evaluation references.

Paper / DOI Dataset & release information

FunkeLab / funlib.evaluate

ERL scoring, including merge, split and omission handling.

Repository ERL source MIT license

AI4Sci fruit-fly reconstruction track

Model definitions, experimental protocols, results and compute measurements. The models are trained on the track data; architectural and protocol differences limit direct comparison with the paper.

Track PR & experiment record Model & evaluation code

October 2026.

Task description and code

Leaderboard

Simulated

Design preview · leaderboard values and research histories are simulated.

Agent search and reference methods ranked together by Segmentation IoU.
#	Method	Type	Segmentation IoU ↑	Neuron recall ↑	Topology agreement ↑	Research time	Run
1		Agent search	0.7646	0.4728	0.6338	76 h
2		Agent search	0.7636	0.4722	0.6330	60 h
3	PATHFINDER (single-checkpoint FFN / SENSE / SHAPE)	Reference	0.7600	0.4700	0.6300	—	—

How to read these leaderboard metrics

These leaderboard values and research histories are simulated design examples. The leaderboard is ranked by Segmentation IoU. The research curves show relative improvement in this metric over PATHFINDER (single-checkpoint FFN / SENSE / SHAPE); only accepted new-best results update the retained trajectory. A positive gain means a better result, with the sign adjusted for metrics where lower is better. The reconstruction task and measured experiment use nERL; the preview IoU, recall and topology columns do not report those experiment results.

Segmentation IoUPrimary metric · Higher is better: Intersection over union between predicted and reference segmentation regions. It measures spatial overlap, not whether neuronal connectivity is preserved.
Neuron recallHigher is better: The share of reference neurons recovered under the evaluation’s matching criteria.
Topology agreementHigher is better: Agreement between reconstructed and reference connectivity under the task’s topology comparison rules. It complements voxel-level segmentation metrics.

AI4AI research progress

Research trajectory

Best retained improvement over the reference baseline.

Runtime–performance frontier

Evaluated candidates and their observed frontier.

How improvement is measured

Relative improvement over PATHFINDER (single-checkpoint FFN / SENSE / SHAPE) in the primary metric; only retained results update the best-so-far curve.

Accepted new-best experiments raise the trajectory. The frontier includes evaluated candidates; trial, partial and stopped experiments are excluded.