Volumetric Overlap

This node computes a set of standard volumetric overlap metrics between corresponding labels in two masks. These measures quantify how well two binary segmentations agree, providing insight into similarity (agreement) and error (disagreement) from both directions. Background voxels (label 0) are ignored.

For each pair of masks Reference ( R ) and Test (T) (or for each label within each mask), the following metrics are returned:

• Dice Coefficient:

  \[ \text{Dice} = \frac{2\,|R \cap T|}{|R| + |T|} \]

  A measure of overlap ranging from 0 (no overlap) to 1 (perfect overlap). Doubly weights the intersection, making it sensitive to small objects.

• Jaccard Index (Intersection over Union):

  \[ \text{Jaccard} = \frac{|R \cap T|}{|R \cup T|} \]

  Also ranges from 0 to 1. Unlike Dice, it treats false positives and false negatives symmetrically in the denominator.

• Volume Similarity:

  \[ \text{Similarity} = 1 - \frac{|\,|R| - |T|\,|}{|R| + |T|} \]

  Reflects how similar the volumes of the two masks are, without regard for spatial alignment. A value of 1 means identical volumes; lower values indicate volume discrepancies.

• False Positive Error (FPE):

  \[ \text{FPE} = \frac{|T \setminus R|}{|T|} \]

  The fraction of voxels in the test mask (T) that are not in the reference mask ( R ). Indicates over-segmentation relative to the reference.

• False Negative Error (FNE):

  \[ \text{FNE} = \frac{|R \setminus T|}{|R|} \]

  The fraction of voxels in the reference mask ( R ) that are missing from the test mask ( R ). Indicates under-segmentation.

Each metric is computed either globally across the entire volume or on a slice-by-slice basis (if Slice-wise is enabled). These measures are suitable for evaluating segmentation algorithms, comparing manual vs. automated labels, and tracking performance across datasets.

Inputs

See also

Keywords: Dice, Jaccard, Similarity, False Positive Error, False Negative Error

Copyright © 2025, Hero Imaging AB