Keywords: Ultrasound, video object detection, guided needle visualization
TL;DR: Comparison of lightweight 2D, 2.5D and 3D detectors for localizing needles in ultrasound videos
Abstract: We investigate video-based deep learning approaches for detecting needle insertions in ultrasound videos. We introduce two efficient and conceptually simple extensions to convert standard 2D object detectors into video object detectors that make use of temporal information from a history of frames. We compare our approaches to a 2D baseline method that makes independent predictions per frame. Given the need to run in real-time on computationally restricted environments, emphasis is placed on low computational complexity.
Paper Type: validation/application paper
Primary Subject Area: Detection and Diagnosis
Secondary Subject Area: Application: Other
Paper Status: original work, not submitted yet
Source Code Url: This short paper introduces conceptually simple extensions to standard 2D object detectors that should be easily reproducible by reading the text. Ability to publish full source code will need to be cleared by legal department. The authors can provide further support to interested readers that wish to replicate results.
Data Set Url: The dataset cannot be released due to privacy and proprietary reasons.
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.