Scholarlink

This paper introduces the initial edition of the Trauma TeleHelper for Operational Medical Procedure Support and Offline Network (Trauma THOMPSON) Challenge. It was organized as a satellite event of the 26th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2023. The challenge contains two tracks and four tasks related to automatic analysis of videos and images about emergency care procedures under resource constrained environments. The three tasks for Track 1, are (1) action recognition; (2) action anticipation; and (3) activity recognition. For Track 2, the only task was visual question answering. The videos were recorded by a team of doctors from the first-person view and annotated by medical professionals. The data were split into 70% for training and 30% for testing. For Task 1, the best method using VideoSwin with Swin-S and ThreeCrop achieved a Top 1 accuracy of 35.27%. For Task 2, the best method using VideoSwin with Swin-S and CenterCrop achieved Top 1 accuracy of 23.67%. No submission was received for Task 3. For the VQA task, the best method relying on MCAN-large with VinVL and FQCA obtained an accuracy of 74.35%.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Overview of the Trauma THOMPSON Challenge at MICCAI 2023

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yupeng Zhuo,
  • Andrew W. Kirkpatrick,
  • Kyle Couperus,
  • Oanh Tran,
  • Jonah Beck,
  • DeAnna DeVane,
  • Ross Candelore,
  • Jessica McKee,
  • Christopher Colombo,
  • Chad Gorbatkin,
  • Eleanor Birch,
  • Bradley Duerstock,
  • Juan Wachs

摘要

This paper introduces the initial edition of the Trauma TeleHelper for Operational Medical Procedure Support and Offline Network (Trauma THOMPSON) Challenge. It was organized as a satellite event of the 26th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2023. The challenge contains two tracks and four tasks related to automatic analysis of videos and images about emergency care procedures under resource constrained environments. The three tasks for Track 1, are (1) action recognition; (2) action anticipation; and (3) activity recognition. For Track 2, the only task was visual question answering. The videos were recorded by a team of doctors from the first-person view and annotated by medical professionals. The data were split into 70% for training and 30% for testing. For Task 1, the best method using VideoSwin with Swin-S and ThreeCrop achieved a Top 1 accuracy of 35.27%. For Task 2, the best method using VideoSwin with Swin-S and CenterCrop achieved Top 1 accuracy of 23.67%. No submission was received for Task 3. For the VQA task, the best method relying on MCAN-large with VinVL and FQCA obtained an accuracy of 74.35%.