Hausdorff Distance

The Hausdorff Distance is defined as the maximum of the directed distances between two non-empty point sets.

• Directed Hausdorff: h(A, B) = max_{a∈A} min_{b∈B} ||a - b||
• Symmetric Hausdorff: H(A, B) = max( h(A, B), h(B, A) )

This symmetric version ensures the metric is commutative and captures the worst-case deviation in both directions.

Because the standard Hausdorff Distance is catastrophically sensitive to a single outlier pixel, the 95th percentile variant is the clinical standard.

• Mechanism: Instead of the absolute maximum, it calculates the distance at the 95th percentile of all boundary distances.
• Robustness: This eliminates the influence of spurious annotation noise while still capturing clinically significant boundary errors.
• Usage: HD95 is mandated in benchmarks like the Medical Segmentation Decathlon.

Understanding the asymmetry is critical for debugging segmentation models.

• False Negative Error: A high h(Prediction, Ground Truth) indicates the model missed a true boundary (under-segmentation).
• False Positive Error: A high h(Ground Truth, Prediction) indicates the model hallucinated a boundary extending beyond the true anatomy (over-segmentation).
• Symmetric H: Reports the worst of these two errors, providing a single conservative metric.

Hausdorff Distance complements Dice Score and IoU by measuring boundary shape rather than area overlap.

• Dice/IoU: Measure the ratio of overlapping area. A model can have a high Dice score (e.g., 0.95) but a catastrophic localized boundary spike.
• Hausdorff: Detects these critical spikes. A single deep protrusion or missed corner will result in a high HD value, alerting clinicians to potentially dangerous segmentation errors.
• Clinical Relevance: In radiotherapy, a 5mm boundary error on a spinal cord OAR is clinically unacceptable, even if the Dice score is high.

Calculating the exact Hausdorff Distance requires computing all pairwise distances, which is computationally expensive for high-resolution 3D volumes.

• Brute Force: O(n*m) complexity where n and m are the number of boundary points.
• Optimization: Early exit algorithms and distance transform maps reduce this to near-linear time.
• Implementation: Libraries like SimpleITK and MONAI provide optimized, GPU-accelerated implementations using distance transforms to calculate HD95 efficiently on 3D medical images.

The Average Hausdorff Distance (AVD) provides a balanced view of boundary error by averaging all distances instead of taking the maximum.

• Formula: The mean of the directed distances from every point in A to the nearest point in B.
• Interpretation: AVD is less sensitive to outliers than standard HD but more sensitive to overall boundary smoothness than HD95.
• Use Case: Often reported alongside HD95 to provide a complete picture of both average and worst-case boundary fidelity.

A statistical measure of spatial overlap between a predicted segmentation mask and the ground truth annotation. Calculated as twice the intersection divided by the sum of the two areas, it ranges from 0 (no overlap) to 1 (perfect overlap). Unlike Hausdorff Distance, the Dice Score is insensitive to boundary shape—two segmentations can have a high Dice Score while exhibiting significant local boundary deviations. It is the most widely reported metric in medical imaging literature.

• Formula: 2|A ∩ B| / (|A| + |B|)
• Strength: Robust to class imbalance
• Weakness: Ignores boundary topology and worst-case errors

A metric quantifying the ratio of the overlapping area to the total union area between a predicted segmentation and the ground truth. More conservative than the Dice Score, IoU penalizes false positives and false negatives more severely. It serves as the primary evaluation standard for object detection and instance segmentation tasks.

• Formula: |A ∩ B| / |A ∪ B|
• Relationship to Dice: Dice = 2 * IoU / (1 + IoU)
• Use Case: Preferable when both over-segmentation and under-segmentation must be penalized equally

The mean of all Euclidean distances from each point on the predicted boundary to the nearest point on the ground truth boundary, and vice versa. While Hausdorff Distance captures the worst-case error, ASD captures the typical boundary deviation. Reporting both metrics together provides a complete picture of boundary quality.

• Symmetric ASD: Averages distances in both directions (predicted→ground truth and ground truth→predicted)
• Unit: Millimeters in physical space after isotropic resampling
• Clinical Relevance: Critical for radiotherapy where systematic boundary offsets affect dose distribution

A robust variant of the Hausdorff Distance that uses the 95th percentile of surface distances instead of the maximum. This eliminates sensitivity to single outlier points caused by annotation errors or imaging artifacts while still capturing clinically significant boundary deviations. HD95 is the recommended standard in most medical segmentation challenges.

• Calculation: Sort all surface distances, take the value at the 95th percentile
• Advantage: Stable across annotator variability
• Reporting: Always report alongside standard HD for transparency

A metric that evaluates boundary agreement by measuring the fraction of surface points within a tolerance margin (typically 1-2 mm) of the reference surface. SDO bridges the gap between volume-based metrics like Dice and distance-based metrics like Hausdorff Distance, providing a clinically interpretable measure of boundary accuracy.

• Tolerance: Defined based on clinical requirements and voxel spacing
• Interpretation: SDO of 0.9 at 2 mm means 90% of predicted boundary is within 2 mm of ground truth
• Application: Increasingly adopted in radiotherapy contouring evaluations

A post-processing algorithm that identifies and labels isolated contiguous regions in a binary segmentation mask. Used to remove small spurious predictions (false positive islands) or separate touching objects before computing boundary metrics like Hausdorff Distance. Without this step, scattered single-voxel predictions can dramatically inflate HD values.

• Typical Filter: Remove components smaller than a clinically relevant volume threshold
• Implementation: Available in scikit-image, MONAI, and SimpleITK
• Impact: Can reduce HD by orders of magnitude by eliminating noise