ivtmetrics

The ivtmetrics library provides a Python implementation of metrics for benchmarking surgical action triplet detection and recognition.

Features at a glance

The following are available with ivtmetrics:

1. Recognition Evaluation: Provides AP metrics to measure the performance of a model on action triplet recognition.
2. Detection Evaluation: Supports Intersection over Union distances measure of the triplet localization with respect to the instruments.
3. Flexible Analysis: (1) Supports for switching between frame-wise to video-wise averaging of the AP. (2) Supports disentangle prediction and obtained filtered performance for the various components of the triplets as well as their association performances at various levels.

Installation

Install via PyPi

To install ivtmetrics use pip

pip install ivtmetrics

Install via Conda

conda install -c nwoye ivtmetrics

Python 3.5-3.9 and numpy and scikit-learn are required.

Metrics

The metrics have been aligned with what is reported by CholecT50 benchmark. ivtmetrics can be imported in the following way:

import ivtmetrics

The metrics implement both recognition and detection evaluation. The metrics internally implement a disentangle function to help filter the triplet components as well as triplet different levels of association.

Recognition Metrics

Recognition ivtmetrics can be used in the following ways:

metric = ivtmetrics.Recognition(num_class)

This takes an argument num_class which is default to 100

The following function are possible with the Recognition class:

args:

• args component can be any of the following ('i', 'v', 't', 'iv', 'it','ivt') to compute performance for (instrument, verb, target, instrument-verb, instrument-target, instrument-verb-target) respectively. default is 'ivt' for triplets.
• args ignore_null (optional, default=False): to ignore null triplet classes in the evaluation. This option is enabled in CholecTriplet2021 challenge.
• the output is a dict with keys("AP", "mAP") for per-class and mean AP respectively.

Example usage

import ivtmetrics
recognize = ivtmetrics.Recognition(num_class=100)
network = MyModel(...) # your model here 
# training
for epoch in number-of-epochs:
  recognize.reset()
  for images, labels in dataloader(...): # your data loader
    predictions = network(image)
    recognize.update(labels, predictions)
  results_i = recognize.compute_AP('i')
  print("instrument per class AP", results_i["AP"])
  print("instrument mean AP", results_i["mAP"])
  results_ivt = recognize.compute_AP('ivt')
  print("triplet mean AP", results_ivt["mAP"])

# evaluation
recognize.reset_global()
for video in videos:
  for images, labels in dataloader(video, ..): # your data loader
    predictions = network(image)
    recognize.update(labels, predictions)
  recognize.video_end()
    
results_i = recognize.compute_video_AP('i')
print("instrument per class AP", results_i["AP"])
print("instrument mean AP", results_i["mAP"])

results_it = recognize.compute_video_AP('it')
print("instrument-target mean AP", results_it["mAP"])

results_ivt = recognize.compute_video_AP('ivt')
print("triplet mean AP", results_ivt["mAP"])

Any nan value in results is for classes with no occurrence in the data sample.

Detection Metrics

Detection ivtmetrics can be used in the following ways:

metric = ivtmetrics.Detection(num_class, num_tool, threshold=0.5)

This takes an argument num_class which is default to 100 and num_tool which is default to 6

The following function are possible with the Detection class:

args:

1. list of list format: [[tripletID, toolID, toolProbs, x, y, w, h], [tripletID, toolID, toolProbs, x, y, w, h], ...], where:

   • tripletID = triplet unique identity
   • toolID = instrument unique identity
   • toolProbs = instrument detection confidence
   • x = bounding box x1 coordiante
   • y = bounding box y1 coordinate
   • w = width of the box
   • h = height of the box
   • The [x,y,w,h] are scaled between 0..1

2. list of dict format: [{"triplet":tripletID, "instrument":[toolID, toolProbs, x, y, w, h]}, {"triplet":tripletID, "instrument":[toolID, toolProbs, x, y, w, h]}, ...].

3. format args describes the input format with either of the values ("list", "dict")

4. component can be any of the following ('i', 'v', 't', 'iv', 'it','ivt') to compute performance for (instrument, verb, target, instrument-verb, instrument-target, instrument-verb-target) respectively, default is 'ivt' for triplets.<

• the output is a dict with keys("AP", "mAP", "Rec", "mRec", "Pre", "mPre") for per-class AP, mean AP, per-class Recall, mean Recall, per-class Precision and mean Precision respectively.

Example usage

import ivtmetrics
detect = ivtmetrics.Detection(num_class=100)

network = MyModel(...) # your model here

# training

format = "list"
for epoch in number of epochs:
  for images, labels in dataloader(...): # your data loader
    predictions = network(image)
    labels, predictions = formatYourLabels(labels, predictions)
    detect.update(labels, predictions, format=format)
      
  results_i = detect.compute_AP('i')
  print("instrument per class AP", results_i["AP"])
  print("instrument mean AP", results_i["mAP"])
    
  results_ivt = detect.compute_AP('ivt')
  print("triplet mean AP", results_ivt["mAP"])
  detect.reset()


# evaluation

format = "dict"
for video in videos:
  for images, labels in dataloader(video, ..): # your data loader
    predictions = network(image)
    labels, predictions = formatYourLabels(labels, predictions)
    detect.update(labels, predictions, format=format)
  detect.video_end()
    
results_ivt = detect.compute_video_AP('ivt')
print("triplet mean AP", results_ivt["mAP"])
print("triplet mean recall", results_ivt["mRec"])
print("triplet mean precision", results_ivt["mPre"])

Any nan value in results is for classes with no occurrence in the data sample.

Disentangle

Although, the Detection() and Recognition() classes uses the Disentangle() internally, this function can still be used independently for component filtering in the following ways:

filter = ivtmetrics.Disentangle()

Afterwards, each of the component's predictions/labels can be filtered from the main triplet's predictions/labels as follows:

i_labels = filter.extract(inputs=ivt_labels, component="i")
v_preds  = filter.extract(inputs=ivt_preds, component="v")
t_preds  = filter.extract(inputs=ivt_preds, component="t")
iv_labels = filter.extract(inputs=ivt_labels, component="iv")
it_labels = filter.extract(inputs=ivt_labels, component="it")

Triplet Association Scores (TAS)

This assesses the quality of bounding box - triplet ids association. using the following metrics:

• LM: localize and match percent: percentage of triplets localized at threshold (θ) and matched with correct triplet ids.
• PLM: partially localize and match: percentage of triplets matched with correct triplet ids but localization overlap is less than θ.
• IDS: identity switch: percentage of triplets localized at θ but with swapped ids within the frame.
• IDS: identity miss: percentage of triplets localized at θ but with incorrect ids (not swapped).
• MIL: missed localization: percentage of triplets matched with correct triplet ids with no matching localization bounding boxes.
• FP: remaining false positives: remaining false alarms after all other scores has been considered.
• FN: remaining false negatives: remaining missed predictions after all other scores has been considered.

The TAS metrics are automatically computed within the Detection class of ivtmetrics. The results are accessed using the TAS metrics acronymns as keys, such as:

import ivtmetrics
detect = ivtmetrics.Detection(num_class=100)
"""
  after a series of detect.update() call    
"""
results_ivt = detect.compute_video_AP('ivt')
print("triplet matched and localized", results_ivt["ml"])
print("triplet identity switchd", results_ivt["ids"])
print("triplet missed localization", results_ivt["mil"])

Docker

coming soon ..

Citation

If you use this metrics in your project or research, please consider citing the associated publication:

@article{nwoye2022data,
  title={Data Splits and Metrics for Benchmarking Methods on Surgical Action Triplet Datasets},
  author={Nwoye, Chinedu Innocent and Padoy, Nicolas},
  journal={arXiv preprint arXiv:2204.05235},
  year={2022}
}

References

1. Nwoye, C. I., Yu, T., Gonzalez, C., Seeliger, B., Mascagni, P., Mutter, D., ... & Padoy, N. (2021). Rendezvous: Attention Mechanisms for the Recognition of Surgical Action Triplets in Endoscopic Videos. arXiv preprint arXiv:2109.03223.
2. Nwoye, C. I., Gonzalez, C., Yu, T., Mascagni, P., Mutter, D., Marescaux, J., & Padoy, N. (2020, October). Recognition of instrument-tissue interactions in endoscopic videos via action triplets. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 364-374). Springer, Cham.
3. http://camma.u-strasbg.fr/datasets
4. https://cholectriplet2022.grand-challenge.org
5. https://cholectriplet2021.grand-challenge.org

License

BSD 2-Clause License

Copyright (c) 2022, Research Group CAMMA
All rights reserved.

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:

1. Redistributions of source code must retain the above copyright notice, this
   list of conditions and the following disclaimer.

2. Redistributions in binary form must reproduce the above copyright notice,
   this list of conditions and the following disclaimer in the documentation
   and/or other materials provided with the distribution.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.```