itk_evaluate

Evaluate segmentation predictions against ground truth by calculating comprehensive metrics and volume statistics.

Usage

itk_evaluate <gt_folder> <pred_folder> <save_folder> [options]

Parameters

• gt_folder: Folder containing ground truth masks (.mha, .nii, .nii.gz, .nrrd files)
• pred_folder: Folder containing prediction masks (same format as gt_folder)
• save_folder: Folder to save evaluation results (created if not exists)
• --format {csv,excel}: Output format (default: excel)
• csv: Saves 5 separate CSV files
• excel: Saves 1 Excel file with 5 sheets
• --mp: Enable multiprocessing for faster evaluation
• --workers N: Number of worker processes (default: half of CPU cores)

Features

• Automatic Resampling: Predictions are automatically resampled to match ground truth spacing/size if they differ
• Consistent Orientation: All samples are oriented to LPI for consistent evaluation
• Multi-class Support: Handles multi-class segmentation with per-class metrics
• Volume Statistics: Calculates volumes in cubic millimeters for both GT and predictions
• Multiple Aggregation Views: Provides metrics in 5 different aggregation formats

Metrics Calculated

For each class, the following metrics are computed:

• Dice Coefficient: Measures overlap between GT and prediction (F1-score)
• IoU (Jaccard): Intersection over Union
• F-score: Same as Dice for segmentation tasks
• Recall (Sensitivity): True positive rate
• Precision: Positive predictive value
• Accuracy: Overall pixel-wise accuracy (global metric)

Volume Statistics

For each class in each sample:

• Volume_GT: Volume in cubic millimeters calculated from ground truth
• Volume_Pred: Volume in cubic millimeters calculated from prediction

Volumes are computed as: voxel_count × (spacing_x × spacing_y × spacing_z)

Output Tables

The tool generates 5 different views of the evaluation results:

1. Per-Class Sample-Averaged (PerClass_SampleAvg)

Mean metrics for each class across all samples.

metric	class_0	class_1	class_2	...
dice	0.95	0.92	0.88	...
iou	0.91	0.85	0.79	...
recall	0.94	0.91	0.87	...

2. Per-Sample Per-Class (PerSample_PerClass)

Detailed metrics for each sample and each class (most granular view).

sample	class_0_dice	class_0_iou	class_1_dice	...	accuracy
case001	0.96	0.92	0.94	...	0.95
case002	0.94	0.89	0.90	...	0.93

3. Per-Sample Class-Averaged (PerSample_ClassAvg)

Mean metric across all classes for each sample.

sample	dice	iou	fscore	recall	precision	accuracy
case001	0.95	0.91	0.95	0.94	0.96	0.95
case002	0.92	0.85	0.92	0.91	0.93	0.93

4. Volume_GT

Volume in cubic millimeters for each class in each sample (from ground truth).

sample	class_0	class_1	class_2	...
case001	1250.5	3420.8	890.2	...
case002	1180.3	3350.1	910.5	...

5. Volume_Pred

Volume in cubic millimeters for each class in each sample (from prediction).

sample	class_0	class_1	class_2	...
case001	1245.2	3380.5	885.1	...
case002	1175.8	3320.3	915.2	...