14:00
Gynecology/Perinatology
Chair: Massimo Mischi
14:00
15 mins
|
Removing electromyogram from non-invasive fetal electrocardiogram using deep learning
Swedha Sankaranarayanan, Rik Vullings, Beatrijs van der Hout-van der Jagt
Abstract: In obstetric practice, cardiotocography (CTG) is the golden standard method to monitor the fetus during labor. CTG is the simultaneous acquisition of fetal heart rate (fHR) and uterine activity. Unfortunately, the interpretation of CTG has a low specificity and consequently leads to unnecessary operative deliveries [1]. Non-invasive fetal electrocardiography (fECG) can be of added value, as it can capture additional information to improve specificity. During labor, the presence of uterine contractions poses an additional risk to fetal health. Monitoring fetal health in this period is therefore extremely important, but at the same time challenging due to strong interferences (e.g. electromyogram (EMG)) in the fECG signals that arise during labor.
Removal of uterine EMG, which can be considered as structured noise, has not yet been explicitly researched. In this abstract, we propose a U-net-based neural network architecture that can learn the structured noise in the spectrum of a typical multi-channel fECG recording by exploiting the relatively narrow bandwidth of fECG (i.e. 15-45 Hz) as compared to EMG (i.e. 0-200 Hz) [2]. Multi-channel pre-processed fECG recordings, obtained from 94 intrapartum patients, are windowed into 30s segments. The segments that contain uterine EMG interferences are automatically selected based on signal statistics per patient recording. The selected segments are filtered using a band-stop filter to remove the fECG and are passed as inputs to the U-net. The neural network learns the missing frequency components of the EMG based on the information and structure present in the rest of its spectrum. Our preliminary results show that such a reconstruction of structured noise is possible. This approach might improve fECG extraction in the presence of uterine contractions.
[1] Z. Alfirevic et al., “Continuous cardiotocography (CTG) as a form of electronic fetal monitoring (EFM) for fetal assessment during labour,” Cochrane Database Syst. Rev., vol. 3, 2006, Art. no. CD006066.
[2] R. Vullings, “Non-invasive fetal electrocardiogram : analysis and interpretation,” Phd Thesis 1 (Research TU/e / Graduation TU/e), Technische Universiteit Eindhoven, 2010. doi: 10.6100/IR692881.
|
14:15
15 mins
|
A detailed 1D model of the feto-placental hemodynamics to investigate hypertensive disorders of pregnancy
Pascalle Wijntjes
Abstract: Hypertensive disorders of pregnancy (HDP), affect a significant number of pregnancies, bringing maternal and fetal health at risk. For better understanding and for improving medical care, the cause of this disorder is increasingly studied. One hypothesis is that HDP patients have an increased placental resistance resulting in insufficient placental perfusion effecting fetal growth.
The hypothesis on increased placental resistance can be tested via 1D modelling of the placental hemodynamics. 1D wave-propagation models can provide crucial insights into pathologies by explaining their influence on clinically measurable Doppler waveform characteristics at the umbilical arteries (UA). Therefore, the aim of this study is to develop a 1D model of the feto-placental circulation and test the hypothesis.
A volume-filling algorithm is employed to generate a realistic placental vasculature, imposing constraints on branching parameters. Murray's law is applied to establish vessel diameter distributions that reflect physiology. The modelled pathological cases corresponding to HDP, have a reduced vessel diameter and incorporate local placental infarcts.
A 1D model, previously developed by Kroon et al., was refined and adapted for this application. At the inlet of the UA, a realistic, model-based, pressure is imposed, whereas at the placental outlets a constant pressure of 20 mmHg is prescribed.
The resulting vasculature mimics real placental statistics and consists of almost 100k 1D elements through which the hemodynamics are calculated. It includes 53 villus trees and 23 generations from UA to capillary. Using a terminal radius at capillary level with a chosen power for Murrays law yields UA diameters of approximately 4 mm. These numbers are all in line with existing literature on healthy placental geometries.
By imposing a realistic pressure of 40 mmHg with a pulse of 8 mmHg, a flow wave form that corresponds to Doppler velocity measurements of the UA is established. The resulting flow is on average 125 mL/min showing good agreement with literature on the mean flow in the umbilical veins. Also, the flow for the pathological UA is reduced.
In summary, our modelling approach adequately mimics feto-placental arterial behaviour and show reduced perfusion for HDP scenarios. Future work will study more pathologies and placental growth.
|
14:30
15 mins
|
The Design Of A 3D Printed Training Phantom Specifically Tailored To Enhance The Training Of Caesarean Sections In Low – And Middle-Income Countries
Dina Khalid Hassan Abubakr, Marjolein Wolbers, Roos Oosting
Abstract: Iatrogenic urological injuries resulting from cesarean sections are significant complications that can lead
to severe patient morbidity and mortality. We have conducted systematic review that investigates the
incidence, causative factors, and preventive measures related to these injuries during cesarean deliveries,
with a particular focus on low- and middle-income countries (LMICs). Analyzing 31 studies, we found
that the incidence of bladder and ureteral injuries is notably higher in LMICs, primarily due to the
involvement of less experienced surgical personnel.
Key injurious procedures identified include the reflection of the bladder flap, which often occurs during
the laparotomy stage, and inadequate mobilization of the bladder during the hysterotomy. The review
highlights that bladder injuries commonly occur at the dome, while ureteral injuries are more frequent
on the left ureter. Such injuries can lead to prolonged hospital stays, additional surgeries, and significant
psychological distress for patients, underscoring the critical need for improved surgical training.
In response to these findings, we are designing a training phantom specifically tailored to enhance
surgical education in LMICs. This phantom will be 3D printed and utilize silicon-based materials to
accurately simulate the skin and the anatomical layers of the abdomen, providing a realistic training
experience for clinicians. Importantly, we intend to develop this training tool at a reasonable cost,
ensuring its accessibility and availability in resource-limited settings. By equipping surgeons with better
training tools, we hope to reduce the incidence of iatrogenic injuries, improve patient outcomes, and
ultimately enhance the overall quality of cesarean sections.
We will present the findings from our review that we plan to publish in the upcoming months, and the
first iteration of a prototype of the training phantom during the BME conference. This marks the start of
a four-year Ph.D. project that will focus on this important topic.
|
14:45
15 mins
|
A novel tool for non-invasive monitoring of uterine electrical activity during pregnancy via electrohysterography
Giulia Acquaviva, Alessandra Galli, Elisabetta Peri, Massimo Mischi
Abstract: Research question: Uterine electrical activity analysis is crucial to monitor pregnancy progression and uterine contractions. This practice is performed using various monitoring tools, among which electrohysterography (EHG) is a promising technique due to its non-invasiveness and reliability [1]. However, there is a lack of methods able to distinguish between uterine contractions and basal uterine activity. Then, this study aims at developing a Machine Learning (ML) model for automatic contraction detection from EHG signals, while simultaneously optimizing the electrode configuration to maximize the classification performance.
Methods: Data is retrieved from the Icelandic 16-electrode EHG Database [2] and includes 6 women in labor. EHG signals are pre-processed by removing the powerline noise at 50 Hz and filtered in [0.1, 1] Hz (via Butterworth filters), then down-sampled to 20 Hz. Signals are subsequently segmented using TOCO as reference and 8 features are extracted for each segment. These features include nonlinear (entropy, nonlinear correlation coefficient, time reversibility) and frequency-related features (RMS, mean frequency). The features that best discriminate between contractions and non-contractions are identified using the Kolmogorov-Smirnov statistical test. This procedure is repeated for 8 electrode configurations (monopolar and bipolar) to determine the most performant one.
Preliminary results: Until now, EHG analysis has been conducted on the monopolar configuration. For each feature, statistical tests were performed between the distributions of contraction and non-contraction segments. The results showed significant differences across all the features. This outcome is particularly evident in the case of RMS and entropy (p < 0.001). In agreement with physiology, RMS values result higher for contractions because the energy of the signal is greater when contractions occur. Entropy values are expected to be lower for regular time-series; during uterine contractions, the increased coordination among cells reduces the irregularity of signals and leads to a decrease of entropy, as proven by our results.
Conclusions: Initial findings indicate the potential of specific features for detecting uterine contractions. Next, we will optimize the electrode configuration accounting for signal propagation directionality and interelectrode distance. Furthermore, a ML classifier will be trained and validated for automatic contraction detection, incorporating new data to enhance both accuracy and reliability.
|
15:00
15 mins
|
Differentiating Adenomyotic from Healthy Uteri based on Heterogeneity in Contrast-Ultrasound Dispersion Imaging
Ferenc I. Kandi, Eva de Bock, Catarina Dinis Fernandes, Simona Turco, Lynda Juffermans, Judith Huirne, Massimo Mischi
Abstract: Adenomyosis is a prevalent benign uterine disorder, often causing heavy menstruations and subfertility. Its diffuse appearance on B-mode ultrasound and MRI complicates its diagnosis. Microvasculature visualization by dynamic contrast-enhanced ultrasound (CEUS) is being explored as a diagnostic method. Featured as driving mechanism in adenomyosis, angiogenesis is the formation of dense/tortuous microvasculature, which can be imaged with CEUS by characterization of the transport kinetics of ultrasound contrast agents (UCAs). In adenomyotic uteri, angiogenesis develops heterogeneously throughout the myometrium. Our aim is to identify objective metrics to aid clinicians in diagnosing adenomyosis based on parametric maps extracted from CEUS, characterizing the underlying microvasculature.
The UCA transport kinetics are modeled as a convective dispersion process and the dispersion of UCA is assessed by analysis of the spectral coherence (ρ) between neighboring time-intensity curves [1]. The 90th percentile of ρ values are assessed to indicate angiogenesis (low dispersion). To assess heterogeneity, the range (R), interquartile range (IQR) and standard deviation (STD) of ρ values are investigated. For the analysis, 18 adenomyosis patients and 25 healthy controls were scanned using a Samsung HERA W10 platform with an EV2-10A endocavital probe operating in contrast-specific mode (UCA bolus injection, power modulation, gain 30 dB, dynamic range 45 dB, 3.2 MHz) at low mechanical index (MI = 0.1). The R, IQR and STD values were calculated per subject, the adenomyotic and healthy groups were compared using Wilcoxon Rank Sum tests. The study protocol was approved by the AUMC ethical review board, all participants signed informed consent.
The R (p = 6.3E-6), IQR (p = 0.0033) and STD (p = 0.0030) of ρ were significantly different when comparing adenomyotic and healthy uteri. The 90th percentile did not conclusively differentiate between adenomyotic and healthy uteri. This is because adenomyosis develops gradually, and healthy uteri may already exhibit early signs of the condition, leading to similarly low dispersion values.
In conclusion, the quantification of heterogeneity through spectral coherence R, IQR and STD can significantly differentiate adenomyotic from healthy uteri.
[1] Mischi et al. Angiogenesis imaging by spatiotemporal analysis of ultrasound contrast agent dispersion kinetics, IEEE T-UFFC, 59(4), 621–629, 2012.
|
15:15
15 mins
|
Transmission Sequence Design for a Pregnancy Monitoring Annular CMUT
Maria Jose Almario Escorcia, Rob van Schaijk, Willem-Jan de Wijs, Richard Lopata, Hans-Martin Schwab
Abstract: High-risk pregnancies can lead to unwanted outcomes for the fetus and mother. For timely intervention, it is vital to perform close monitoring. Ultrasound provides relevant fetal information but its current image form factors are not typically suited for continuous monitoring. As a foundational development stage toward an accessible and compact ultrasound monitoring device for high-risk pregnancies, an annular capacitive micromachined ultrasonic transducer (CMUT) was designed and manufactured. The array is small and cost-effective but poses imaging challenges, such as poor steerability. This study proposes a point spread function (PSF) based performance analysis of transmission sequences to maximize image quality for this low-resource transducer geometry.
We developed a flexible GPU-accelerated ultrasound simulation framework, used for this analysis. A point scatterer was positioned at increasing depths (z-axis) and different radial positions. Different transmission schemes were confronted. An unfocussed (all elements transmitting at t = 0s) and a focused transmission were used as baseline. These wavefronts were selected as we theorized that the results of other sequences would fall between their outcomes. For the designed sequence, we explored the use of diverging waves with coherent compounding, testing various arrangements of virtual sources. The sequences were compared by analyzing the PSF using three metrics: full-width-half-maximum (FWHM, in mm) of the main lobe of the x-axis PSF profile, at z and y coordinates fixed by the scatterer position; the -6 dB separability of two scatterers on said profile (in mm); and contrast ratio. Additionally, we performed acquisitions of a thin needle submerged in water to mimic a single scatterer.
Simulations using virtual sources following a sunflower pattern lead to the best outcomes among the investigated schemes. The average metrics results of the compounding approach outperform focusing by up to 73%. All metrics show trends that are in accordance, with the contrast ratio adding value to the analysis by accounting for contributions of side lobes. Acquisitions and simulations of the same scenario exhibit comparable behavior. In future work, we will expand the analysis by incorporating additional metrics and explore the use of this method to evaluate new CMUT design concepts.
|
|