Inquiry regarding examples for certain instruments

Dear organizers,

The data description page displays examples of all the instruments expect 2: bipolar dissector and suction irrigator:

• Suction irrigator: It's included in the exception of the rule: if the tool is very large and the clevis does not always show in the field of view (e.g tip-up fenestrated grasper, suction irrigator, stapler, grasping retractor) Does this mean that the middle joint of the suction irrigator is the ground truth and the big white tip is only considered in case the joint is not shown? Could you please provide us with an example as this instrument doesn't have the typical tip/clevis/shaft as the rest of the instruments.

• Bipolar dissector: There are 17 clips with this instrument and the instruments in those clips don't look similar to the robotic ones in the data description page. Most of the instruments in those 17 clips have different shapes and colors, the clips also don't include the UI as the other clips. What is considered ground truth in those clips? and how should we deal with the instruments difference with respect to the rest of the clips? For example permanent cautery hook/spatula has a yellow clevis in most clips but in these 17 clips it has a white joint.

17 clips in question: ['clip_016647', 'clip_016648', 'clip_016649', 'clip_016650', 'clip_016651', 'clip_016652', 'clip_016653', 'clip_016654', 'clip_016655', 'clip_016656', 'clip_016657', 'clip_016658', 'clip_016659', 'clip_016660', 'clip_016661', 'clip_016662', 'clip_016663']

Best regards,

Hi,

1. We have added an example bounding box for the suction irrigator at the bottom of the Data Description page. Please refer to this for the correct bounding box, and feel free to post any follow up questions.

2. We have checked these clips and saw that the tools in question are laparoscopic tools. These clips are a part of the noise in the dataset, which is based on robotic tool install data but can contain laparoscopic instruments in the field of view as well.

Let us know if you have any further questions!

Best, The SurgToolLoc Organizing Committee

1. Thanks for the new example :)

2. Regarding the bipolar dissector, it was labelled only in those clips, which means that there is no representation of this instrument in the dataset for the algorithm to learn its features. How should we deal with this situation in case this instrument happens to be in the test set?

3. I have another question regarding the clevis/tip of some instruments: The second rule states The tool is very large and the clevis does not always show in the field of view (e.g tip-up fenestrated grasper, suction irrigator, stapler, grasping retractor). Does this apply to every instrument, for example cadiere forceps or bipolar forceps? If the clevis of the cadiere forceps is not in the field of view (which is very frequent) the tip should always be predicted?

Best,

1. We are looking into the dataset. It looks like you are correct that these are the only examples in the training dataset, in which case we will remove "bipolar dissector" from the evaluation set so it will not affect algorithm results. Since this only affects 17/25k videos, it will hopefully not affect existing algorithms.
2. This does not apply to every image. If the base of tip is visible, then yes the bounding box should capture this. However, if the base of tip is not visible then the instrument would not be annotated.

2. This does not apply to every image. If the base of tip is visible, then yes the bounding box should capture this. However, if the base of tip is not visible then the instrument would not be annotated.

This means ground truth is clevis in case tip and clevis are appeared, while ground truth "for any instrument" is tip in case only tip is appeared (clevis is not appeared). Is it correct? If so, clevises of these instruments(tip-up fenestrated grasper, suction irrigator, stapler, grasping retractor) are sometimes appeared, but the ground truth bboxes are tip for these instruments?

I am confused now, because I first thought that the instruments were grouped to:

• The ground truth is always the whole tip+clevis: Tip-up fenestrated grasper, suction irrigator, stapler, grasping retractor. Because the data description examples of the grasping retractor and Tip-up fenestrated grasper show a full box on tip+clevis, although the clevis is clearly present in the frame.

• Others: Clevis only (if the clevis doesn't appear, the tip is not annotated)

It would be nice if we could have extra examples for both cases (clevis appearing in the frame, clevis not appearing in the frame) for the instruments.

Best regards,

Hi,

To clarify, the "base of the tip" is the outer bounds of the clevis as well.

For large tools, the ground truth is clevis + tip when visible, and tip when the clevis is not visible.

For other tools, the ground truth is clevis, where the base of the tip is the outer bound of the clevis. I have added an example on the data description page where only the base of the tip is visible and annotated by a bounding box, with a comparison to a bounding box for a tool that is fully visible.

Let me know if this clears up your questions.

Best, SurgToolLoc Organizing Committee

Thank you for the clarification and for the new example.

Dear organizer,

I have additional questions related to annotations.

1. If the tip is hidden by some tissue in the following case(e.g. clip_000370), is the ground truth of monopolar curved scissors None?

2. If tip only (not with clevis) appears in the following case(e.g. clip_000035), is the ground truth None? In this case, cadiere forceps only appears at the tip, so no ground truth?

Best, Atsushi

Hi Atsushi,

1. If the entire annotated portion of the image (usually clevis, see rules for details) is obscured by tissue, it would not have a bounding box annotated. Otherwise, the visible portion would be annotated.

2. Yes, if only the tip is visible there would be no annotation.

I understood it. Thank you for clarifying these questions.

Sincerely, Atsushi

Hi @cperreault,

Based on the conversation above, is there any confirmation regarding the tool "bipolar dissector" not been included in the final metric computation?

Hi,

Yes as stated above bipolar dissector is not included in the evaluation set due to the low amount of training samples.

Dear organizer,

I have a question about the ground truth of the clevis: according to the Data Description page, the ground truth of the cadiere forceps and needle driver does not include the part sandwiched by the shaft (Figure 1, 2). However, if the wrist is folded at 90 degrees and is on the back side of the screen, the clevis may not be visible as shown in Figure 3 (only visible the part sandwiched by the shaft which is excluded in the ground truth in Figure 1, 2). In this case, is there no ground truth?

Figure 1¶

Figure 2¶

Figure 3¶

Best, Atsushi

Hi Atsushi,

The bounding box in this case would be small, but will have a ground truth annotation where the clevis is.

Let me know if this answers your question.