Surgical Action Triplet Detection 2022
Latest News
| 30-11-2022 | CholecTriplet2021 challenge joint paper accepted at Medical Image Analysis journal. Check out the arXiv report. | |
| 18-09-2022 | CholecTriplet2022 challenge results announced. Check out the results and winners. | |
| 18-09-2022 | Personal efforts on the tasks involving localization of the triplets (or triplet's instruments) cannot be published until a joint publication of the challenge methods is released on the arXiv. | |
| 13-04-2022 | CholecT45 is a subset of CholecT50 dataset excluding the challenge test set. CholecT45 has been released under CC BY-NC-SA 4.0 license on April 12, 2022. Starting from this date, it can be used for submissions to conferences, journals and other venues. Download access is through the CAMMA website. | |
| 12-04-2022 | Official splits of the dataset for developing deep learning models is contained in arxiv. | |
| 11-04-2022 | Joint publication of the last edition of the challenge, CholecTriplet2021, is now on the arxiv. | |
Overview
Formalizing surgical activities as triplets of the used instruments, actions performed, and target anatomies acted upon provides a better, comprehensive and fine-grained modeling of surgical activities. Automatic recognition of these triplet activities directly from surgical videos would facilitate the development of intra-operative decision support systems that are more helpful, especially for safety, in the operating room (OR).
Our previous EndoVIS challenge, CholecTriplet2021 (MICCAI 2021), and existing works on surgical action triplet recognition tackles this as a multi-label classification of all possible <instrument, verb, target> combinations. For better clinical utility, real-time modeling of tool-tissue interaction will go beyond determining the presence of these action triplets, to also include estimating their locations in each video frame.
Hence, this challenge extends our previous challenge on action triplet recognition to also include bounding box localization of the regions of action triplets. For the lack of spatially annotated dataset and to exploit large dataset without expensive and tedious annotation effort, this challenge focuses on the development and evaluation of weakly-supervised approaches for bounding box localization of action triplets on CholecT50 dataset.
Participants will develop and compete with algorithms to recognize action triplets as well as localize their region of likelihood in laparoscopic videos without the use of spatial annotation during training. This novel challenge investigates the state-of-the-art on surgical fine-grained activity detection, weak supervised learning of action location, and will strengthen this new promising research direction on surgical fine-grained activity modeling in computer-assisted surgery.
This year's challenge will feature a Colab code blog providing some sample codes for quick start, and T50 slack interactive forum to share ideas and interacts with organizers and co-participants. The teams with the best results will be awarded prizes. We also plan a joint publication with top participants after the challenge. It indeed will be a rewarding experience.
Interesting? Join the challenge now!
Additional Information
- The CholecTriplet2022 sub-challenge is part of the Endoscopic Vision Challenge MICCAI 2022.
- A complete structured description of the accepted 2022 EndoVis challenge design can be found here.
- The dataset for the challenge is CholecT50. The annotation protocol is briefly explained here.
- The challenge is guided by the following EndoVis rules.
Sponsors
Industrial Sponsors:
Academic Sponsors:
