Pymia
When the pipeline is configured to use the pymia backend (backend: "pymia"), segmentation evaluation is
delegated to the robust pymia image analysis library.
References:
Below is an outline of the metrics currently mapped and supported by this project's pymia wrapper, including their
exact configuration keys. It is important to note that not all metrics provided by pymia are necessarily meaningful
or commonly used in the context of segmentation evaluation. The pymia library exposes a broad set of evaluation
measures, some of which are designed for more general machine learning or statistical comparison tasks. Nevertheless,
for completeness and flexibility, the wrapper implemented in AUDIT exposes most of the metrics currently supported by pymia,
allowing users to select those that best fit their specific evaluation needs.
Supported metrics configuration
To extract these metrics, use the exact Attribute name under the metrics section of your configuration file.
Overlap metrics
| Metric Name | Attribute Name (Config) |
|---|---|
| Adjusted Rand Index | ari |
| Area Under Curve (AUC) | auc |
| Cohen Kappa Coefficient | ckc |
| Dice Coefficient | dice |
| Interclass Correlation | ic |
| Jaccard Coefficient (IoU) | jacc |
| Mutual Information | mi |
| Rand Index | rand |
| Surface Dice Overlap | sdo |
| Surface Overlap | so |
| Volume Similarity | vs |
Distance metrics
| Metric Name | Attribute Name (Config) |
|---|---|
| Average Distance | avd |
| Global Consistency Error | gce |
| Hausdorff Distance (100th %ile) | haus |
| Mahalanobis Distance | mahal |
| Probabilistic Distance | prob |
| Variation Of Information | vi |
Classical metrics
| Metric Name | Attribute Name (Config) |
|---|---|
| Accuracy | accu |
| Fallout | fallout |
| F-Measure | fmeas |
| False Negative Count | fn |
| False Negative Rate | fnr |
| False Positive Count | fp |
| Precision (PPV) | prec |
| Prediction Volume | pred_vol |
| Reference Volume | ref_vol |
| Sensitivity (Recall) | sens |
| Specificity | spec |
| True Negative Count | tn |
| True Positive Count | tp |
Excluded regression measures
pymia also implements coefficient of determination (R²), MAE, MSE, RMSE,
and normalized RMSE. AUDIT does not expose them as segmentation metrics. For
binary masks, MAE and MSE collapse to the same pixel error rate, RMSE is only
its square root, normalized RMSE is redundant for a 0–1 range, and R² is
strongly driven by foreground prevalence. Accuracy and the overlap metrics are
more interpretable for this workflow.
Mahalanobis distance remains available because pymia calculates it from the foreground coordinates: it measures centroid displacement relative to the regions' joint covariance. It should nevertheless be interpreted cautiously for very small or geometrically degenerate regions.