Reconstruction Metrics

swclib provides five complementary metrics for comparing a predicted neuron reconstruction against a gold standard. All metrics share the same interface:

result = metric.run(gold_swc_path, pred_swc_path)

gold_swc_path and pred_swc_path are paths to SWC files. The return value is always a plain dict.

---

Overview

Metric	Class	What it measures
SSD	SSDMetric	Mean spatial distance between node sets
Point	PointMetric	Node-level nearest-neighbor precision/recall
Length	LengthMetric	Edge-level path coverage (precision/recall)
Keypoint	KeypointMetric	Detection of branch and leaf points
Fiber	FiberMetric	Path-level (root-to-leaf) matching

---

SSD Metric

The Spatial Distance (SSD) metric resamples both trees to uniform node spacing, then computes the mean distance between each node and its nearest neighbor in the other tree.

Usage

from swclib.metrics.ssd_metric import SSDMetric

metric = SSDMetric(
    min_distance=2.0,       # resampling step (µm)
    scale=(1, 1, 1),        # coordinate scaling
)
result = metric.run("gold.swc", "pred.swc")

Output

Key	Description
sd	Mean bidirectional spatial distance
sd_gt2pred	Mean distance from gold nodes to nearest predicted node
sd_pred2gt	Mean distance from predicted nodes to nearest gold node
ssd	Mean of distances exceeding min_distance
ssd_gt2pred	SSD in gold→pred direction
ssd_pred2gt	SSD in pred→gold direction
ssd_percent	Fraction of nodes with distance > min_distance

---

Length Metric

The Length metric compares reconstructions at the edge level. Each edge (segment between two consecutive nodes) in the gold standard is matched against edges in the predicted tree within a spatial radius.

Usage

from swclib.metrics.length_metric import LengthMetric

metric = LengthMetric(
    radius_threshold=2.0,   # matching radius (µm)
    length_threshold=0.2,   # minimum matched length (µm)
    scale=(1, 1, 1),
    resample_step=2.0,      # resampling step for edge comparison
)
result = metric.run("gold.swc", "pred.swc")

Output

Key	Description
precision	Fraction of predicted length that matches gold
recall	Fraction of gold length that is covered by prediction
f1	Harmonic mean of precision and recall
TP	Matched edge length
FP	Unmatched predicted edge length
FN	Unmatched gold edge length

---

Point Metric

The Point metric compares all resampled nodes by nearest-neighbor distance. It is useful when you want a simple node-level precision/recall score.

Usage

from swclib.metrics.point_metric import PointMetric

metric = PointMetric(
    dist_threshold=4.0,
    scale=(1, 1, 1),
    resample_step=2.0,
)
result = metric.run("gold.swc", "pred.swc")

Output

Key	Description
precision	Fraction of predicted points near gold points
recall	Fraction of gold points covered by prediction
f1	Harmonic mean of precision and recall
MES	Morphology error score-style overlap ratio
TP / FP / FN	Matched / extra / missed points

---

Keypoint Metric

The Keypoint metric evaluates detection of topologically significant points: branch nodes (degree ≥ 3) and leaf nodes (degree 1).

Usage

from swclib.metrics.keypoint_metric import KeypointMetric

metric = KeypointMetric(
    keypoint_types=["branch", "leaf"],
    threshold_dis=5.0,      # matching distance threshold (µm)
    scale=(0.35, 0.35, 1),
    use_category=True,      # report branch and leaf stats separately
    mode="block",           # or "whole"
)
result = metric.run("gold.swc", "pred.swc")

Output (combined mode, use_category=False)

Key	Description
precision	Keypoint detection precision
recall	Keypoint detection recall
f1	F1 score
TP / FP / FN	Matched / extra / missed keypoints

Output (per-category mode, use_category=True)

The dict contains flat per-category fields:

result["branch_precision"]
result["branch_recall"]
result["branch_f1"]

result["leaf_precision"]
result["leaf_recall"]
result["leaf_f1"]

---

Fiber Metric

The Fiber metric is the most comprehensive metric. It extracts all root-to-leaf paths (fibers) from both trees and matches them using Intersection-over-Union (IoU).

A fiber is counted as a True Positive (TP) if its IoU with a matched gold fiber exceeds iou_threshold.

Usage

from swclib.metrics.fiber_metric import FiberMetric

metric = FiberMetric(
    iou_threshold=0.8,          # minimum IoU for a match to count as TP
    dist_threshold=5.0,         # spatial tolerance (µm) for IoU calculation
    dist_sample=1.0,            # resampling step for IoU calculation
    scale=(1, 1, 1),
    only_from_soma=True,        # only evaluate fibers starting at type=1 nodes
    min_fiber_length=20.0,      # ignore fibers shorter than this (µm)
    align_roots=True,           # align tree roots before comparison
    align_roots_thredhold=20.0, # current source spelling; max distance to match roots
    with_direction=False,       # consider fiber direction
    use_category=True,          # separate axon (type=2) and dendrite (type=3/4) stats
)
result = metric.run("gold.swc", "pred.swc")

Output

Key	Description
precision	Fraction of predicted fibers that match a gold fiber
recall	Fraction of gold fibers that are matched by a prediction
f1	Harmonic mean
TP / FP / FN	Fiber-level counts
num_gt	Number of gold fibers
num_pred	Number of predicted fibers
iou_matched	Mean IoU among matched fiber pairs
iou_all	Mean IoU across all gold fibers (unmatched = 0)
ious	Per-fiber IoU values
matches	List of matched (gold_fiber, pred_fiber) pairs
FN_fiber_ids	Node IDs of unmatched gold fibers

With use_category=True, per-category stats appear as flat fields such as axon_precision, axon_recall, axon_f1, dendrite_precision, dendrite_recall, and dendrite_f1.

With return_fibers=True, the actual SwcFiber objects are returned under result["gold_fibers"] and result["pred_fibers"].

Example: inspect unmatched fibers

result = metric.run("gold.swc", "pred.swc", return_fibers=True)

for i, fiber in enumerate(result["gold_fibers"]):
    iou = result["ious"][i]
    if iou < metric.iou_threshold:
        print(f"Unmatched gold fiber, length={fiber.length:.1f}, IoU={iou:.2f}")

---

Batch evaluation with MetricManager

MetricManager runs multiple metrics on a dataset of reconstruction pairs and aggregates results.

from swclib.metrics.manager import MetricManager

manager = MetricManager(
    metric_names=["ssd", "length", "keypoints", "fiber"],
    collect_method="micro",  # aggregate TP/FP/FN across all pairs ("micro")
    scale=(1, 1, 1),
)

# Optionally override per-metric parameters
manager.update_metric_params({
    "fiber": {"iou_threshold": 0.8, "dist_threshold": 5.0},
    "keypoints": {"threshold_dis": 5.0},
})

# Add reconstruction pairs
for gold_path, pred_path in zip(gold_files, pred_files):
    manager.add_data(gold_path, pred_path)

# Aggregate and save
summary = manager.collect(save_path="results.json")
print(summary)

Collect methods

Method	Description
"micro"	Sum all TP/FP/FN across the dataset, then compute final P/R/F1
"macro"	Not implemented in the current source

Micro averaging is typically preferred for imbalanced datasets (neurons with very different fiber counts).

---

Choosing the right metric

Scenario	Recommended metric
Quick sanity check	SSD
Evaluate path tracing completeness	Length
Evaluate topology accuracy	Keypoint
Comprehensive morphology evaluation	Fiber
Full benchmark pipeline	MetricManager with SSD, length, keypoint, and fiber metrics