ECG Waveform Segmentation for P-QRS-T Detection Using Deep Learning Based on Fourier Synchrosqueezed Transform

Nian Zhang; Roni A. Romero Melendez; J'niya Butler; Wagdy H Mahmoud; Lara A Thompson

Download Paper | Permalink

Conference: ASEE Mid-Atlantic Section Spring Conference
Location: George Washington University, District of Columbia
Publication Date: April 19, 2024
Start Date: April 19, 2024
End Date: April 20, 2024
Page Count: 7
DOI: 10.18260/1-2--45747
Permanent URL: https://peer.asee.org/45747
Download Count: 298

Dr. Nian Zhang is a Professor in the Department of Electrical and Computer Engineering at the University of the District of Columbia (UDC), Washington, D.C., USA. Her research interests include machine learning, deep learning, classification, clustering, and optimization.

visit author page

author page

J'niya Butler University of the District of Columbia

biography

Wagdy H Mahmoud P.E. University of the District of Columbia

visit author page

Wagdy H. Mahmoud is a Professor of electrical engineering at the Electrical Engineering Department at UDC. Mahmoud is actively involved in research in the areas of reconfigurable logic, hardware/software co-design of a system on a chip using reconfigurabl

visit author page

author page

Lara A Thompson University of the District of Columbia

Download Paper | Permalink

Abstract

Segmenting brain scans is a valuable technique for visualizing distinct brain structures, and it finds common use in quantifying volumes and shapes to characterize both healthy and afflicted populations. Manual segmentation conducted by clinical experts is regarded as the gold standard, but it's excessively time-consuming and impractical for annotating extensive datasets. Moreover, it necessitates a deep understanding of neuroanatomy and is susceptible to errors and variability in both interrater and intrarater reproducibility.

To mitigate these challenges, machine learning approaches, particularly convolutional neural networks (CNNs), have emerged as promising tools for automating the labeling of substantial clinical datasets. Nevertheless, despite progress in data augmentation and transfer learning, CNNs struggle to generalize effectively to unseen data domains. When applied to brain scans, CNNs exhibit high sensitivity to variations in image resolution and contrast. Even when dealing with MRI data from the same modality, their performance may degrade when handling diverse datasets.

Inspired by the pretrained SynthSeg neural network, we aim to develop a 3-D U-Net based on CNNs for brain MRI segmentation. This algorithm will segment brain scans of varying contrast and resolution without requiring retraining or fine-tuning. It will achieve this versatility through training with synthetic data generated from a generative model conditioned on existing segmentations. Specically, we will employ a domain randomization strategy, fully randomizing the contrast and resolution of the synthetic training data. As a result, our proposed method is expected to effectively segment real scans from a wide array of target domains without necessitating retraining or fine-tuning, making it feasible to analyze extensive and heterogeneous clinical datasets. In addition, we will employ test time augmentation to enhance segmentation accuracy. In essence, augmentation applies random transformations to an image to increase the dataset's variability. We will utilize augmentation prior to network training to expand the training dataset. Test time augmentation will involve applying random transformations to test images, thereby creating multiple versions of each test image. Subsequently, each version of the test image will be presented to the network for prediction. The final segmentation result will be computed as the average prediction across all versions of the test image. This strategy is anticipated to improve segmentation accuracy by mitigating the impact of random errors in individual network predictions.

We will compare the performance of the proposed model with CNNs and Bayesian segmentation on different MRI images datasets. We will also investigate the generalizability of the proposed method by applying it to cardiac MRI and/or CT scans.

Citation
Format

Romero Melendez, R. A., & Zhang, N., & Butler, J., & Mahmoud, W. H., & Thompson, L. A. (2024, April), ECG Waveform Segmentation for P-QRS-T Detection Using Deep Learning Based on Fourier Synchrosqueezed Transform Paper presented at ASEE Mid-Atlantic Section Spring Conference, George Washington University, District of Columbia. 10.18260/1-2--45747

TY - CPAPER
AB - Segmenting brain scans is a valuable technique for visualizing distinct brain structures, and it finds common use in quantifying volumes and shapes to characterize both healthy and afflicted populations. Manual segmentation conducted by clinical experts is regarded as the gold standard, but it&#39;s excessively time-consuming and impractical for annotating extensive datasets. Moreover, it necessitates a deep understanding of neuroanatomy and is susceptible to errors and variability in both interrater and intrarater reproducibility.

To mitigate these challenges, machine learning approaches, particularly convolutional neural networks (CNNs), have emerged as promising tools for automating the labeling of substantial clinical datasets. Nevertheless, despite progress in data augmentation and transfer learning, CNNs struggle to generalize effectively to unseen data domains. When applied to brain scans, CNNs exhibit high sensitivity to variations in image resolution and contrast. Even when dealing with MRI data from the same modality, their performance may degrade when handling diverse datasets.

Inspired by the pretrained SynthSeg neural network, we aim to develop a 3-D U-Net based on CNNs for brain MRI segmentation. This algorithm will segment brain scans of varying contrast and resolution without requiring retraining or fine-tuning. It will achieve this versatility through training with synthetic data generated from a generative model conditioned on existing segmentations. Specically, we will employ a domain randomization strategy, fully randomizing the contrast and resolution of the synthetic training data. As a result, our proposed method is expected to effectively segment real scans from a wide array of target domains without necessitating retraining or fine-tuning, making it feasible to analyze extensive and heterogeneous clinical datasets. In addition, we will employ test time augmentation to enhance segmentation accuracy. In essence, augmentation applies random transformations to an image to increase the dataset&#39;s variability. We will utilize augmentation prior to network training to expand the training dataset. Test time augmentation will involve applying random transformations to test images, thereby creating multiple versions of each test image. Subsequently, each version of the test image will be presented to the network for prediction. The final segmentation result will be computed as the average prediction across all versions of the test image. This strategy is anticipated to improve segmentation accuracy by mitigating the impact of random errors in individual network predictions.

We will compare the performance of the proposed model with CNNs and Bayesian segmentation on different MRI images datasets. We will also investigate the generalizability of the proposed method by applying it to cardiac MRI and/or CT scans.

AU - Roni A. Romero Melendez
AU - Nian Zhang
AU - J'niya Butler
AU - Wagdy H Mahmoud P.E.
AU - Lara A Thompson
CY - George Washington University, District of Columbia
DA - 2024/04/19
PB - ASEE Conferences
TI - ECG Waveform Segmentation for P-QRS-T Detection Using Deep Learning Based on Fourier Synchrosqueezed Transform
UR - https://peer.asee.org/45747
DO - 10.18260/1-2--45747
ER -