MitraClip Device Automated Localization in 3D Transesophageal Echocardiography via Deep Learning

TL;DR

This paper introduces an automated deep learning pipeline for accurate detection and classification of the MitraClip device in 3D transesophageal echocardiography images, improving visualization and procedural guidance.

Contribution

It presents a novel combination of Attention UNet segmentation, DenseNet classification, and CAD-based template registration for MitraClip detection in 3D TEE images.

Findings

Achieved high segmentation accuracy with 0.76 mm surface distance

Classified device configuration with 75% F1-score

Enhanced visualization and quantitative assessment of the device

Abstract

The MitraClip is the most widely percutaneous treatment for mitral regurgitation, typically performed under the real-time guidance of 3D transesophagel echocardiography (TEE). However, artifacts and low image contrast in echocardiography hinder accurate clip visualization. This study presents an automated pipeline for clip detection from 3D TEE images. An Attention UNet was employed to segment the device, while a DenseNet classifier predicted its configuration among ten possible states, ranging from fully closed to fully open. Based on the predicted configuration, a template model derived from computer-aided design (CAD) was automatically registered to refine the segmentation and enable quantitative characterization of the device. The pipeline was trained and validated on 196 3D TEE images acquired using a heart simulator, with ground-truth annotations refined through CAD-based templates. The Attention UNet achieved an average surface distance of 0.76 mm and 95% Hausdorff distance of 2.44 mm for segmentation, while the DenseNet achieved an average weighted F1-score of 0.75 for classification. Post-refinement, segmentation accuracy improved, with average surface distance and 95% Hausdorff distance reduced to 0.75 mm and 2.05 mm, respectively. This pipeline enhanced clip visualization, providing fast and accurate detection with quantitative feedback, potentially improving procedural efficiency and reducing adverse outcomes.