fidoptBrachial plexus in neck ultrasound videos
One frame of a patient's neck ultrasound video, 2D, with the supraclavicular or interscalene brachial plexus boxed by anaesthesiologists and shrunk to its outline by an active contour (AIIMS New Delhi and IIIT Hyderabad).
22 test caseshead and neckimage US2D, distances in pxresearch only
Read before comparing (8)
- Videos. 227 patients, one video each: 196 on an eSaote MyLab One (562 x 542 pixels, 20 frames a second), and 15 each on a Sonosite M-Turbo (512 x 360) and a Butterfly iQ (566 x 534) by other operators, with one demonstration video. The six eSaote videos without a plexus and the demonstration are left out. Frames are H.264 as ffmpeg decodes them, in colour for a few coloured marks; no pixel size is given, so distances are in pixels.
- Frame. Every frame has a mask, empty where no plexus shows; the anaesthesiologist boxed a median of four frames a video by hand, the tracker the rest. The case is the middle hand-boxed frame, whose boxes a person drew, though its outline is still the active contour's; for a video boxed by hand once or twice it is the first of those frames (18 of the 41 cases), frame 0 of the video for 15.
- Truth. Coarse by design: the boxes cover the plexus as a whole, not its roots, and stay clear of the subclavian artery and the pleura, and the paper scores detection more than outlines. Graded semi-automatic, so ranked apart.
- Levels. eSaote videos are supraclavicular (157) or interscalene (33), one label for both; the Sonosite and Butterfly videos are supraclavicular, six Sonosite videos with a needle in view.
- Overlap. No other dataset of the corpus holds brachial plexus ultrasound; the Kaggle 2016 nerve set is another, closed collection. SonoCorpus's MD5 manifest finds none of these videos in its other 52 datasets.
- Outside numbers. Test cases are drawn inside SonoCorpus's test list, as KvasirSEG's inside FCBFormer's test split: SonoBase trained on its 154 training videos and was validated on its 22 validation videos, none a test case. The authors' own models were cross-validated over all eSaote videos and have seen the test patients.
- Start. On the 22 test frames the plexus covers 4.7% of a frame on average: the whole frame scores a Dice of 0.09, Otsu's bright and dark pixels 0.15 and 0.06. TotalSegmentator has no ultrasound task.
- Tolerance. 3 pixels: the truth grown or thinned by one pixel keeps a Dice of 0.97 and a surface Dice of 1.0 at 3 pixels.
No method has a result on every case of this dataset yet.
Case by case
Click a score to look at that result: the image, the truth in green and the method's map in red. Underlined scores have their image; for the others you see the case without the result.
About the data
- Test cases
- 22 test patients drawn with seed 2026 among the 44 videos of SonoCorpus's test list, stratified by scanner, so that SonoBase has seen none of them; 19 development patients from its 176 training and validation videos, in the seeded order. One video a patient and one frame a video: the middle one of the frames boxed by hand.
- Truth
- contours by a program, verified by people the provider's ac_masks: an expert anaesthesiologist boxed the plexus on seed frames, a tracker carried the boxes to the other frames, each frame approved or redrawn by the annotator, and the union of the boxes was shrunk by a morphological geodesic active contour whose parameters anaesthesiologists chose video by video (Tyagi et al., IROS 2024, section III.A)
- Source
- doi.org/10.7910/DVN/EVLM8B
github.com/Regional-US/brachial_plexus/tree/ac4bf778db
arxiv.org/abs/2308.03717
doi.org/10.1109/IROS58592.2024.10801467
doi.org/10.5281/zenodo.22770825 - Licence
- CC BY-NC-SA 4.0 (the Harvard Dataverse deposit)
19 cases for tuning, never tested
eSaote_3380, eSaote_4374, eSaote_4971, eSaote_1104, Butterfly_b937, eSaote_8157, eSaote_8505, eSaote_4981, eSaote_9291, eSaote_6185, eSaote_4605, eSaote_8224, eSaote_5088, eSaote_2416, eSaote_4206, eSaote_2940, Butterfly_b473, eSaote_5247, Sonosite_s187