LabelBenchmark¶
LabelBenchmark compares predicted and reference label images at foreground
pixel level. It reports object counts and simple foreground agreement metrics.
Inputs are predicted_label_image and reference_label_image. Outputs include
predicted and reference label counts, true-positive, false-positive, and
false-negative pixel counts, and foreground_iou.
Use it for small segmentation benchmark demos and smoke tests. It does not
perform instance matching; future measurement backlog tools should add object
matching, Dice, IoU, and panoptic metrics. Shape mismatches raise a ValueError.
Dependencies and Core Libraries¶
BioImageFlow core APIs, imageio, and NumPy.
Assumptions¶
Both inputs are aligned 2D label images where foreground is label > 0.
Minimal Example¶
from bioimageflow_core import Arguments
from bioimageflow_measurement_tools import LabelBenchmark
LabelBenchmark().process_row(
Arguments(predicted_label_image="pred.tif", reference_label_image="gt.tif")
)
Expected Results¶
The tool returns foreground true-positive, false-positive, false-negative pixel counts and foreground IoU.
Failure Modes¶
Shape mismatches raise ValueError; missing or unsupported files fail through
imageio.