Development and Analysis of a Pipeline for Cardiac Ultrasound Simulation for Deep Learning Segmentation Methods
No Thumbnail Available
Date
2024
Journal Title
Journal ISSN
Volume Title
Publisher
The Eurographics Association
Abstract
Accurate and efficient segmentation of anatomical structures in medical images, e.g. ultrasound images, is crucial for diagnosis. Deep Learning methods can provide automatic reproducible segmentation, and simulation of medical images with their intrinsic ground truth could help to develop and tune these methods. We introduce a simulation pipeline for the example of mitral valve segmentation in Transesophageal Echocardiography (TEE) images including different valve opening states. As anatomical ground truth, we used a CT based patient phantom with simulated mitral valve closure. For each region within the phantom, scatter intensities and reflections between tissue boundaries were set, and ultrasound images were simulated with incorporation of attenuation and noise. To further improve realism of the simulated images a speckle reduction filter was used. The adjustments applied to improve realism were assessed by testing the segmentation performance (including Dice score) of a deep learning method trained on real TEE data. The initial Dice score for the simulation was 31 %. This value increased with image postprocessing (37 %), exclusion of surrounding cardiac structures (45 %) and the combination of both (46 %). In comparison, the initial Dices score for real TEE was 72 %. On both simulated and real TEE images, the deep learning method performed better on fully closed valve states (42 % and 77 %) than on fully open valves (27 % and 66 %). This work introduced a novel pipeline for the realistic simulation of TEE images with different valve opening states. Our analysis demonstrated feasibility of the proposed pipeline and highlighted the importance of accurate and dynamic valve phantoms, comprehensive simulations and specific post-processing for the simulation of realistic TEE images. In the future, with further improvements of the simulation, we will evaluate the pipeline for the training of Deep Learning methods on simulated data for the application on real data.
Description
CCS Concepts: Computing methodologies → Modeling and simulation; Human-centered computing → Visualization
@inproceedings{10.2312:vcbm.20241189,
booktitle = {Eurographics Workshop on Visual Computing for Biology and Medicine},
editor = {Garrison, Laura and Jönsson, Daniel},
title = {{Development and Analysis of a Pipeline for Cardiac Ultrasound Simulation for Deep Learning Segmentation Methods}},
author = {Bauer, Marcel and Manini, Chiara and Klemmer, Stefan and Meyer, Tom and Ivantsits, Matthias and Walczak, Lars and Hennemuth, Anja and Tzschätzsch, Heiko},
year = {2024},
publisher = {The Eurographics Association},
ISSN = {2070-5786},
ISBN = {978-3-03868-244-8},
DOI = {10.2312/vcbm.20241189}
}