A package to compute different segmentation metrics for 2D/3D medical images.
Project description
Segmentaion Metrics Package
This is a simple package to compute different metrics for Medical image segmentation(images with suffix .mhd
, .mha
, .nii
, .nii.gz
or .nrrd
), and write them to csv file.
BTW, if you need the support for more suffix, just let me know by creating new issues
Summary
To assess the segmentation performance, there are several different methods. Two main methods are volume-based metrics and distance-based metrics.
Metrics included
This library computes the following performance metrics for segmentation:
Voxel based metrics
- Dice (F-1)
- Jaccard
- Precision
- Recall
- False positive rate
- False negtive rate
- Volume similarity
The equations for these metrics can be seen in the wikipedia.
Surface Distance based metrics (with spacing as default)
- Hausdorff distance
- Hausdorff distance 95% percentile
- Mean (Average) surface distance
- Median surface distance
- Std surface distance
Note: These metrics are symmetric, which means the distance from A to B is the same as the distance from B to A. More detailed explanication of these surface distance based metrics could be found here.
Installation
$ pip install seg-metrics
Getting started
Tutorial is at the Colab.
API reference is available at Documentation
Examples could be found below.
Usage
At first, import the package:
import seg_metrics.seg_metrics as sg
Evaluate two batches of images with same filenames from two different folders
labels = [0, 4, 5 ,6 ,7 , 8]
gdth_path = 'data/gdth' # this folder saves a batch of ground truth images
pred_path = 'data/pred' # this folder saves the same number of prediction images
csv_file = 'metrics.csv' # results will be saved to this file and prented on terminal as well. If not set, results
# will only be shown on terminal.
metrics = sg.write_metrics(labels=labels[1:], # exclude background
gdth_path=gdth_path,
pred_path=pred_path,
csv_file=csv_file)
print(metrics) # a list of dictionaries which includes the metrics for each pair of image.
After runing the above codes, you can get a list of dictionaries metrics
which contains all the metrics. Also you can find a .csv
file containing all metrics in the same directory. If the csv_file
is not given, the metrics results will not be saved to disk.
Evaluate two images
labels = [0, 4, 5 ,6 ,7 , 8]
gdth_file = 'data/gdth.mhd' # ground truth image full path
pred_file = 'data/pred.mhd' # prediction image full path
csv_file = 'metrics.csv'
metrics = sg.write_metrics(labels=labels[1:], # exclude background
gdth_path=gdth_file,
pred_path=pred_file,
csv_file=csv_file)
After runing the above codes, you can get a dictionary metrics
which contains all the metrics. Also you can find a .csv
file containing all metrics in the same directory.
Note:
- When evaluating one image, the returned
metrics
is a dictionary. - When evaluating a batch of images, the returned
metrics
is a list of dictionaries.
Evaluate two images with specific metrics
labels = [0, 4, 5 ,6 ,7 , 8]
gdth_file = 'data/gdth.mhd'
pred_file = 'data/pred.mhd'
csv_file = 'metrics.csv'
metrics = sg.write_metrics(labels=labels[1:], # exclude background if needed
gdth_path=gdth_file,
pred_path=pred_file,
csv_file=csv_file,
metrics=['dice', 'hd'])
# for only one metric
metrics = sg.write_metrics(labels=labels[1:], # exclude background if needed
gdth_path=gdth_file,
pred_path=pred_file,
csv_file=csv_file,
metrics='msd')
By passing the following parameters to select specific metrics.
- dice: Dice (F-1)
- jaccard: Jaccard
- precision: Precision
- recall: Recall
- fpr: False positive rate
- fnr: False negtive rate
- vs: Volume similarity
- hd: Hausdorff distance
- hd95: Hausdorff distance 95% percentile
- msd: Mean (Average) surface distance
- mdsd: Median surface distance
- stdsd: Std surface distance
For example:
labels = [1]
gdth_file = 'data/gdth.mhd'
pred_file = 'data/pred.mhd'
csv_file = 'metrics.csv'
metrics = sg.write_metrics(labels, gdth_file, pred_file, csv_file, metrics=['dice', 'hd95'])
dice = metrics['dice']
hd95 = metrics['hd95']
Evaluate two images in memory instead of disk
Note:
- The two images must be both numpy.ndarray or SimpleITK.Image.
- Input arguments are different. Please use
gdth_img
andpred_img
instead ofgdth_path
andpred_path
. - If evaluating
numpy.ndarray
, the defaultspacing
for all dimensions would be1.0
for distance based metrics. - If you want to evaluate
numpy.ndarray
with specific spacing, pass a sequence with the length of image dimension asspacing
.
labels = [0, 1, 2]
gdth_img = np.array([[0,0,1],
[0,1,2]])
pred_img = np.array([[0,0,1],
[0,2,2]])
csv_file = 'metrics.csv'
spacing = [1, 2]
metrics = sg.write_metrics(labels=labels[1:], # exclude background if needed
gdth_img=gdth_img,
pred_img=pred_img,
csv_file=csv_file,
spacing=spacing,
metrics=['dice', 'hd'])
# for only one metrics
metrics = sg.write_metrics(labels=labels[1:], # exclude background if needed
gdth_img=gdth_img,
pred_img=pred_img,
csv_file=csv_file,
spacing=spacing,
metrics='msd')
About the calculation of surface distance
The default surface distance is calculated based on fullyConnected border. To change the default connected type,
you can set argument fullyConnected
as False
as follows.
metrics = sg.write_metrics(labels=[1,2,3],
gdth_img=gdth_img,
pred_img=pred_img,
csv_file=csv_file,
fully_connected=False)
In 2D image, fullyconnected means 8 neighbor points, while faceconnected means 4 neighbor points. In 3D image, fullyconnected means 26 neighbor points, while faceconnected means 6 neighbor points.
How to obtain more metrics? like "False omission rate" or "Accuracy"?
A great number of different metrics, like "False omission rate" or "Accuracy", could be derived from some the confusion matrics. To calculate more metrics or design custom metrics, use TPTNFPFN=True
to return the number of voxels/pixels of true positive (TP), true negative (TN), false positive (FP), false negative (FN) predictions. For example,
metrics = sg.write_metrics(
gdth_img=gdth_img,
pred_img=pred_img,
TPTNFPFN=True)
tp, tn, fp, fn = metrics['TP'], metrics['TN'], metrics['FP'], metrics['FN']
false_omission_rate = fn/(fn+tn)
accuracy = (tp + tn)/(tp + tn + fp + fn)
Comparision with medpy
medpy
also provide functions to calculate metrics for medical images. But seg-metrics
has several advantages.
- Faster.
seg-metrics
is 10 times faster calculating distance based metrics. This jupyter notebook could reproduce the results. - More convenient.
seg-metrics
can calculate all different metrics in once in one function whilemedpy
needs to call different functions multiple times which cost more time and code. - More Powerful.
seg-metrics
can calculate multi-label segmentation metrics and save results to.csv
file in good manner, butmedpy
only provides binary segmentation metrics. Comparision can be found in this jupyter notebook.
If this repository helps you in anyway, show your love ❤️ by putting a ⭐ on this project. I would also appreciate it if you cite the package in your publication. (Note: This package is NOT approved for clinical use and is intended for research use only. )
Citation
If you use this software anywhere we would appreciate if you cite the following articles:
Jia, Jingnan, Marius Staring, and Berend C. Stoel. "seg-metrics: a Python package to compute segmentation metrics." medRxiv (2024): 2024-02.
@article{jia2024seg,
title={seg-metrics: a Python package to compute segmentation metrics},
author={Jia, Jingnan and Staring, Marius and Stoel, Berend C},
journal={medRxiv},
pages={2024--02},
year={2024},
publisher={Cold Spring Harbor Laboratory Press}
}
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for seg_metrics-1.2.8-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 9dc5c1a752f275e41d470d8c422564e3e24a83b0975899758044481d970cded9 |
|
MD5 | 98f0f54281ad4d35ddc5894903e89143 |
|
BLAKE2b-256 | f8d904eb212a0f0d1d31d717516f64358d60a256cc9d7b831cbc9b678cfff895 |