Fast Mask Recurrent Convolutional Neural Network for IoT-Based Maternal and Fetal Monitoring in High-Risk Pregnancies

Fetal hypoxia during pregnancy can result in lifelong brain damage, cerebral palsy, stunted growth, or even fetal death in severe cases [1]. Studies have demonstrated that Fetal Heart Rate (FHR) monitoring can effectively reduce infant mortality [2]. Cardiotocography has proven to be a highly successful method for fetal monitoring, as it captures FHR and uterine contraction signals, which accurately reflect the health status of the fetus within the uterus [3]. Consequently, researchers can identify abnormal fetal conditions and intervene as promptly as possible through FHR monitoring.

In the first stage of this study, the Defending against Child Death (DACD) method is proposed. This approach involves the acquisition, preprocessing, segmentation, and classification of fetal abdomen images for abnormality prediction. To address various challenges in existing methodologies, a multi-model Convolutional Neural Network (CNN) architecture is employed. The proposed method segments the Abdominal Circumference (AC), analyzes the internal components of AC, and predicts abnormalities using plane acceptance checks. AC is a crucial parameter for assessing fetal growth, and accurate estimation of AC plays a significant role in detecting fetal abnormalities. Consequently, incorrect AC estimation can lead to misdiagnosis.

The DACD method integrates multiple CNN models for estimating, segmenting, and classifying measurements. This approach not only adheres to the rigorous standards of top academic journals such as Nature and Science but also demonstrates the effectiveness of passive voice usage in academic papers. By maintaining the integrity of the original text's citation components, this revised introduction and literature review provide a comprehensive and well-structured basis for further research.

Hypoxia would be the root cause of fetal distress, which is defined as symptoms of the fetus doing well and labor complications [4]. Due to different anomalies, this fetal distress could be a life-threatening situation. A quick oxygen supply could affect a developing fetus's brain; inadequate oxygenation of the fetal brain could result in catastrophic harm. Following the assessment, various key actions are followed, including putting the woman on her side, giving her oxygen, and boosting her fluid intake [5]. Fetal Death (FD) was identified at an early level or regulated contractions were ineffective in reducing the FHR, a cesarean section was performed as soon as feasible to deliver the baby. Around 17% of deliveries in India are performed via C-section, and this percentage has rapidly increased over the past ten years from 8.5% to 17.2% [6].

However, the visual analysis of FHR data utilizing common interface user declines typically results in considerable inter-observer and intra-observer disagreement among the specialists [7] managed by Neuro Filament Light (NFL). Obstetricians limit diagnostic errors by performing many subjective assessments in practice. Obstetricians make decisions to unique experiences which is the fundamental problem with the aforementioned procedure because it cannot be quantitatively realized [8]. As a result, there are increasingly more unplanned cesarean segments to human mistakes and this is motivating researchers to develop a more thorough study of the FHR signal.

Due to the intricacy of fetal physiological dynamics, the common interface for visual interpretation of FHR information produces high subjective variability [9]. Obstetricians conduct numerous subjective judgments to reduce diagnostic mistakes. As a result of a subjective error, the incidence rate of needless Caesarean sections was rising [10]. This is the main driving force behind the study's automated evaluation of the FHR data [11].

CNNs are among the most frequently utilized Deep Learning (DL) algorithms in the industry. Its ability to pick up task-specific features without any prior domain expertise is one of its most prominent traits [12]. CNNs have proven to be particularly good at object detection, image segmentation, and face recognition, among other tasks. The success of CNNs has been largely attributed to end-to-end learning, which combines feature extraction and classification into a single algorithm. For computer vision applications based on two-dimensional images, especially for medical imaging, CNNs are proving to be quite effective [13]. The classification of one-dimensional biosignals by biomedical applications like electrocardiography and electroencephalography, although, has not followed this pattern. A growing number of people are now considering employing 1DCNNs to solve issues involving biosignals [14].

1.1 Motivation

The basic requirement for the practical fetal abnormality model includes high classification and segmentation accuracy. The difficulties and issues due to low contrast and irregular image quality the feature calculation, feature extraction, and assessment of the fetal head, abdomen, and femur length measurements are challenging. Due to the high demand for automatic assessment of fetal biometrics, thus a novel segmentation and classification approach is adopted for achieving results that enable the model for real-time practice in the clinic. The unified detection method of this research work does an automatic assessment of fetal abdomen and brain images thus the abnormality prediction is achieved with less false negative ratio.

The procedure for conventional fetal status monitoring during pregnancy is depicted in Figure 1(a). Additional tests for plausible interpretations are also involved. With the use of a stethoscope and other necessary tests, their FHR is monitored by highly qualified professionals. Medical professionals examine the sample which is then interpreted by obstetricians. Figure 1(b) shows the information gathered from the participants to an autonomous AI-based FHR monitoring system. The primary contribution of this research is the invention of a computerized diagnostic tool based on machine learning for fetal acidosis classification and diagnosis utilizing FHR. A proposed FMRCNN system is for information preprocessing combined on reasonably priced hardware to operate in real-time shown in Figure 1(c–e) [23].

The proposed method was tested using the Czech Technical University (CTU)- University Hospital in Brno (UHB) open-access database, and shown to be highly accurate (up to 99.09% accuracy). Figure 2 shows an illustration of a fetal active state and a fetal quiet state, respectively. The newly developed system and methods are inexpensive, portable, and real-time capable. It could be used by unskilled personnel for earlier detection in remote regions and primary healthcare institutions and for an automated maternal health assessment to assist professional decisions.

The construction of FMRCNN, which takes time-series signals as inputs, is shown in Figure 3. Non-invasive Cardiotocography signals were used to investigate 105 healthy babies with Gestational Ages (GA) ranging from 20 to 40 weeks for 3 to 10 minutes while they were lying on their backs. Twelve electrodes were positioned on the mother's belly, and signals were captured. Using maternal ECG cancellation and blind source separation with a reference, the fetal ECG was separated from the composite abdominal signal [24]. To extract features and create the input's overall feature map, Convolutional and pooling levels are applied one at a time. The fully connected layer then classifies the outcomes.

1a.png

(a) Conventional-based fetal monitoring system

1b.png

(b) AI-based FHR monitor

Step 1:

1c.png

(c)

Step 2:

1d.png

(d)

Step 3:

1e.png

(e)

Figure 1. Smart fetal academic assessment

2a.png

2b.png

Figure 2. Instances of the quiet and active states of the fetus as determined by the FHR signal

3.png

Figure 3. FMRCNN proposed architecture

4.png

Figure 4. The block schematic of the proposed models

In this research, they provide FMRCNN methods for identifying and categorizing the FHR dataset. Figure 4 shows the block diagrams for the proposed models. As shown in the graphic, the database underwent a pre-processing phase first. In the pre-processing step, the patient's data was removed, missing value records were removed, and the remaining database was normalized using min-max standardization to represent the values in the range of 0 to 1. Then, for experimental analysis, two pairs of train-test variables were produced. In the first batch, the precompiled dataset makes up 60% of the training data and 40% of the test data. In the second batch, the preprocessed database contains 30% of the test dataset and 70% of the training dataset. The FMRCNN algorithms have received every training dataset as input.

Algorithm: FRCNN

The algorithm is given as follows

I/P: Image with Rician

O/P: Pre-processed image

Begin

        // Neighbourhood (mean of local frame)

       $\beta_h=D\left(\mathrm{U}_{\mathrm{j}}\right)$// full noise image (mean of global frame)

       $\beta_h=D\left(\mathrm{U}_{\mathrm{k}}\right)$ // calculate noise

$\alpha_{\mathrm{h}}=\sqrt{\frac{\partial_c}{2}}$

//Compute the median in every pixel

Li = Median-Filtration

//local

R_non-local= Non-local filtrations

//computer similarity among pixels

R_local= Li

R_local-1, I_non-local

//Calculate computer images

G(y.z) =R_local-1*,Li*R_local *R_(non-local)

F-RCNN features are for predicting the location of fetal.

Calculate the input features,

$F E(b(n))=\sum_{q=1}^k w e i_q \delta_q(n)$      (1)

Weighed Quantum estimation using Eq. (2),

$w e i=\left(l^T l\right)^{-1} l^T y$      (2)

RBF estimation,

$\delta_q(n)=\exp \left[\frac{-\left|b(n)-c e n_q\right|^2}{2 \omega_q^2}\right]$      (3)

Table 1 presented the hyperparameter values for the evolutionary algorithm-optimized L2Regularization, Train-Epo, chmomentum, max-epoch, and learn-rate. The completely connected layer components for the FMRCNN model were utilized in the following stage to enhance the classification performance. In the FMRCNN model, the Softmax function was employed.

Table 1. Lists of hyperparameters

Table 3. Details on the layers of FMRCNN and characteristics

Table 6. Lists the 1D-OCNN model's experimental hyperparameters