Lesion evaluation

Lesion evaluation and components

exacta.lesions treats connected components as lesions and matches predicted lesions to target lesions. Matching conventions affect detection counts and segmentation quality, so connectivity and matching are explicit required arguments.

Matching	Behavior
greedy	One-to-one, best qualifying score first
hungarian	Maximum-total-score one-to-one assignment among qualifying pairs; requires SciPy
merge	Allows additional predicted pieces when their union improves a matched target's score
overlap	Many-to-many matching after target dilation, following BraTS-style evaluation

For 3D data, connectivity 1 means faces (6 neighbors), 2 includes edges (18), and 3 includes corners (26). Choose score="iou" or "dice" for thresholded matching; the default threshold comparison is inclusive, with strict=True for an exclusive comparison.

import numpy as np
from exacta import components, lesions

target = np.zeros((1, 12, 12, 12), dtype=np.int64)
target[:, 2:5, 2:5, 2:5] = 1
prediction = np.roll(target, 1, axis=1)
scores = lesions.evaluate(prediction, target, num_classes=2, include_background=False,
                          connectivity=3, matching="greedy", score="iou", threshold=0.5,
                          spacing=(1.0, 1.0, 1.0))
print(scores.tp, scores.fp, scores.fn, scores.f1)
print(scores.sq_iou, scores.pq_iou, scores.lesion_dice, scores.lesion_hd95)

labels, count = components.label(prediction[0] > 0, connectivity=3)
largest = components.largest(prediction[0] > 0, connectivity=3)
cleaned = components.remove_small(prediction[0] > 0, 5, connectivity=3)

tp, fp, fn, precision, recall, F1, SQ/PQ and lesion scores are per case and selected class. lesion_hd95 requires spacing; without it, that field is NaN. missed_distance controls penalties for missed lesions and false positives. min_target_volume excludes target lesions at or below the threshold, in physical volume units when spacing is supplied (for example, mm³ in 3D), or voxel counts otherwise. Predictions matched to excluded lesions are not counted as false positives; merely touching a lesion need not satisfy the chosen matching rule.

Use lesions.brats_regions and lesions.brats2023(..., challenge="GLI", spacing=...) for the implemented BraTS 2023 convention. Check the dataset's label encoding and challenge settings against the source documentation before applying it to another challenge year.