Weakly-supervised learning for catheter segmentation in 3D frustum ultrasound

Area of Science:

• Medical Imaging
• Computational Imaging
• Interventional Cardiology

Background:

• Accurate catheter segmentation in 3D ultrasound (US) is critical for guiding cardiac interventions.
• Current convolutional neural network (CNN) based methods face challenges with high computational costs and large data sizes, hindering real-time applications.

Purpose of the Study:

• To develop an efficient and accurate catheter segmentation method for 3D ultrasound.
• To address the limitations of traditional Cartesian US data by utilizing Frustum US.
• To overcome the difficulties in annotating irregular Frustum US images through a weakly supervised learning framework.

Main Methods:

• Proposed a novel approach performing catheter segmentation in Frustum US, which has a smaller data volume than Cartesian US.
• Developed a weakly supervised learning framework requiring only bounding-box annotations.
• Generated voxel-level labels using class activation maps combined with line filtering, iteratively updated during training.

Main Results:

• Catheter segmentation in Frustum US achieved a processing time of 0.25 seconds per volume, demonstrating significantly improved efficiency compared to Cartesian US.
• The proposed method achieved better accuracy in catheter segmentation.
• Validated the effectiveness of the weakly supervised learning approach for Frustum US data.

Conclusions:

• Segmentation in Frustum US offers a more efficient alternative to Cartesian US for catheter tracking in 3D ultrasound.
• The proposed weakly supervised learning framework effectively addresses the annotation challenges associated with Frustum US.
• The developed method shows promise for real-time ultrasound-guided cardiac interventions.