Baseline/acceleration/deceleration determination of fetal heart rate signals using a novel ensemble LCResU-Net. (15th May 2023)

• Record Type: Journal Article
• Authors: Liu, Mujun; Zeng, Rongdan; Xiao, Yahui; Bai, Jieyun; Liu, Jia; Zheng, Zheng; Lu, Yaosheng
• Is Part Of: Expert systems with applications. Volume 218(2023)
• Publication Date: 2023-05-15
• Subjects: Baseline -- Acceleration -- Deceleration -- Fetal heart rate -- Electronic fetal monitoring -- Convolutional neural networks; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2023.119610
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: Objective: According to guidelines on electronic fetal heart rate (FHR) monitoring, visual FHR interpretation by obstetricians depends on the recognition of patterns (mainly including baseline, accelerations and decelerations (A/D)). Computer-assisted FHR analysis can effectively reduce obstetricians' inconsistency and improve fetal diagnosis efficiency. Precisely detecting patterns is one of the primary challenges of current computer-aided analysis. This study aims to propose an ensemble U-Net (augmented with residual, layer, and channel shortcut connections) (ELCResU-Net) with minimal post-processing to address the problem. Methods: In this study, A/D are firstly detected with the proposed ELCResU-Net in the FHR signal. After removing detected A/D, this signal is filtered using a progressive trimming procedure based on long-term and short-term frames that consider the continuity and stability of the baseline. Finally, the remaining signal is used to compute the baseline, and thereby A/D of FHR signals are determined. The proposed ELCResU-Net is composed of four LCResU-Nets with kernel sizes of 21, 31, 61 and 81, respectively. The LCResU-Net architecture consists of an encoder block and a corresponding decoder block followed by a point-wise classification layer to construct the 1D segmentation map of A/D from the input FHR signal. In the novel LCRes block of ELCResU-Net, three shortcut connections are used to dynamically prune unimportant channel/layer/block to improve the determination performance. Results: The proposed model is trained on the open-access training data set of the Catholic University of Lille France (CULF-DB) and then tested on its test data set and the independent database of the Jinan University (JNU-DB). Experimental results demonstrate that the proposed method achieves the accelerations' F1-score of 78.82 %, the decelerations' F1-score of 79.10 %, the baseline difference of 2.61 bpm, the synthetic inconsistency coefficient of 49.99 % and the morphological analysis discordance index of 3.89 % in CULF-DB, which are the best performance ever achieved. On the JNU-DB dataset, the performance of the proposed method is also superior to most competing algorithms. Conclusion: The proposed ELCResU-Net for baseline/A/D determination of FHR signals achieves a high level of performance in morphological analysis.