Biomedical Engineering: Applications, Basis and Communications

An accurate assessment of the morphological characteristics of the Optic Disc (OD) is essential for the diagnosis of various retinal disorders. It is necessary to segment the OD precisely to detect structural OD changes associated with visual field loss. Although deep learning models are effective for this task, they require extensive labeled datasets, which can be time-consuming and costly. Furthermore, fundus images have multi-scale features, making segmentation challenging. In this study, we present a semi-supervised and transfer learning approach for OD segmentation. Our approach utilizes an improved Dilated Convolutional AutoEncoder (DCAE) and a pre-trained modified U-Net to segment the OD. The DCAE segments the OD using feature similarity from unlabeled images in the Messidor dataset and saves the learned weights. Transfer learning is then applied to reuse the model weights in the U-Net, accelerating training on small datasets such as Drions-DB and Drishti-GS. The network architecture was modified by increasing the layers from 8 to 128 and halving the feature map length and width. To address the multi-scale challenge without inflating the model parameters, we introduce the Dilated Hierarchical Feature Extraction Module (DHFEM), a convolutional module capable of achieving multi-scale feature extraction without increasing model parameters. Additionally, DHFEM incorporates convolutional layers with varying receptive fields, further enhancing the network ability to extract features across multiple scales. Our OD segmentation method outperforms existing algorithms with reduced parameter quantities of 0.4 M. The mean Intersection over Union (mIoU) values are 0.9383 and 0.9629 and inference times of 45 ms and 40 ms for the Drions-DB and Drishti-GS datasets, respectively.

Colorectal cancer (CRC), ranking third most prevalent cancer type, can be diagnosed with the detection of polyps in the colon and rectum through endoscopic procedures facilitating prompt treatment. During visualization of gastrointestinal tract by the physician, there is high probability of miss rates and reviewing of the images is laborious. Automatic segmentation and detection are enabled with the convolutional neural networks (CNN). We segmented the polyps from the wireless capsule endoscopy images of Kvasir dataset using various CNN models. We have presented nine optimizers for each architecture and evaluated the performance parameters. The optimizers were graded based on the performance metrics in order to provide an insight for the researchers on the selection of optimizer and architecture. On comparison of the performance metrics of the pretrained and U-net-based architecture, the Adaptive Moment Estimation (ADAM) and Root Mean Squared Propagation (RMSPROP) optimizers received the highest score of 43 in the ranking, DiffGrad and Nesterov-accelerated Adaptive Moment Estimation (NADAM) ranked second with the score of 13, the Adaptive Delta (ADADELTA) ranked third with a score of 2, whereas Stochastic Gradient Descent (SGD), Adaptive Gradient Descent (ADAGRAD), and Adaptive Max (ADAMAX) optimizers performed least in the evaluation. Based on the deep learning application, the optimizer employed varies by considering computational speed, memory and computational time. This preliminary research provides the necessary key information for consideration in the development of an architecture with utilization of an optimizer.

The primary objective of this study is to segment the uterine fibroids (leiomyoma) from the ultrasound images of the uterus through semantic segmentation, followed by second-order statistical feature extraction using the Gray-level Co-occurrence Matrix (GLCM). The next objective of the study is to compare the performance of the state-of-the-art method namely Vision Transformer (ViT) with three different machine learning (ML) classifiers such as the Support Vector Machine (SVM), Logistic Regression (LR) and $k$-Nearest Neighbor ($k$-NN) to classify the images into uterine fibroid and normal. The dataset consists of 50 ultrasound images of uterine fibroids and 50 normal images. Then the images are segmented using region-growing-based semantic segmentation followed by feature extraction and classification using the ML and deep learning (DL) classifiers. Among the ML classifiers, SVM produced a good accuracy of 93.1% compared to the other classifiers. ViT produced an excellent classification accuracy of 97.5%. Hence, ViT outperformed compared to the ML classifiers in uterine fibroid detection. These findings have important implications for clinical practice, as they could help physicians to diagnose and treat uterine fibroids more effectively.

Electroencephalography (EEG) is a common clinical method of recording the electrical activity of the brain. EEG can record High-Frequency Oscillations (>80 HZ), which carry appropriate information regarding Epilepsy. High-Frequency Oscillations (HFO) serve as a potential biomarker for Epileptogenesis. EEG signals are often prone to artifact corruptions, which mislead the clinicians by the incorrect signal interpretations. Therefore, automatic artifact removal approach is a key phase in all the Brain-Computer Interface (BCI) applications. In this work, the automatic artifact identification and removal strategy without consuming any supplementary reference channel using two different approaches is developed and discussed. A proficient novel Modified Online Bi-Conjugate Gradient-based Independent Component Analysis (MOBICA) is developed. An efficient threshold-based peak detection and removal strategy, Sparsity-based Artifact Removal Technique (SART) is constructed, where Principle Component Analysis (PCA) is replaced with Singular Value Decomposition (SVD) of the K-SVD algorithm. Both the proposed models are evaluated on two different datasets like CHB-MIT and SRM scalp data recordings. Both the MOBICA and SART algorithms removed the artifactual component parting the intact EEG source component. Finally, the performance of the proposed agenda is compared with the conventional approaches. Our MOBICA and SART algorithms remove the artifactual component parting the intact EEG source component. Empirical results of SART on CHB-MIT and SRM databases of 52 EEG recordings outperform MOBICA maintaining least Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and high Signal to Artifact Ratio (SAR), Mutual Information (MI), and Correlation Coefficient (CC). The proposed strategy vows to be a promising solution for artifact removal in the clinical use of EEG signals and in BCI applications.

Breast cancer detection is a highly fatal disease and is normally detected considering histopathological images (HPIs). However, the complexity of the HPIs makes it challenging to detect breast cancer accurately. Further, manual detection is highly time-consuming and subjective and depends on the experience of the medical professionals. To overcome these issues, an effective deep learning (DL) method for detecting breast cancer from HPIs is proposed. Here, the proposed approach is realized using various processes, such as pre-processing, blood cell segmentation, feature extraction, and classification. Segmentation is accomplished using the SegAN, and classification is performed using the deep convolutional neural network (DCNN). Both networks are trained using the proposed invasive water Ebola optimization (IWEO) algorithm. The efficiency of breast cancer detection is improved by using various features, such as shape features, histogram of gradients (HOG) and local gradient patterns (LGP). Further, the IWEO-DCNN is inspected for its dominance by considering measures, such as accuracy, test negative rate (TNR) and test positive rate (TPR), and the experimental results show that the IWEO-DCNN attained a maximal accuracy of 0.963, TNR of 0.963 and TPR of 0.950.

Retinal vasculature feature extraction plays a critical role in the diagnosis and treatment of systemic conditions, particularly in the cases of diabetic retinopathy (DR). This research introduces an algorithm that utilizes segmented blood vessels in retinal images to identify and differentiate five stages of DR, including mild, moderate, severe and proliferative. The algorithm effectively extracts retinal blood vessels by integrating morphological operators and matched filters, yielding a more precise output. The algorithm’s performance is evaluated using the database IDRiD, demonstrating precision and sensitivity scores comparable to those of a trained observer. A box-counting method was incorporated to measure the fractal dimension (FD) of DR-segmented vessel images at various stages to enhance the accuracy of DR staging. The FD analysis was applied to both thick and thin segments of the blood vessels, enabling the assessment of accuracy, sensitivity and specificity. The results indicate that the algorithm successfully identifies the different stages of DR with an accuracy of 93.65% for the mild stage, of 93.33% for the moderate and severe stages and of 92.71% for proliferative DR compared to the images without DR. The study reveals that the variation in FD between the thick and thin vessel components can be an effective biomarker for identifying the different stages of DR, contributing to a better understanding of disease progression. By combining morphological operators, matched filters and fractal dimension analysis, this research presents a promising approach for specialists involved in diagnosing and treating DR, eventually leading to improved patient care and consequences.

The World Health Organization (WHO) estimates that over 15 million infants are born before the entire period of pregnancy. Over a million neonatal deaths occurred in 2015 as a result of preterm delivery, which is the prime cause for most deaths among children under the age of five. Although preterm birth has no hereditary link, only 10% of preterm deliveries in high-income settings result in deaths, compared to a mortality rate of up to 90% in low-income nations. When preterm cases are discovered, in addition to enhanced care provided at home for the expecting mother, the growing foetus may also benefit from medication, hospitalization for the duration of the pregnancy, or both. Low-income countries also struggle with a lack of access to a comprehensive healthcare system, which makes it difficult to proactively detect premature births. Machine learning techniques have a great potential to improve this situation by providing an intelligent framework for the detection of critical situations and consequently alerting the individual in cases of anomalies. This will require further follow-up with medical specialists. In this work, we investigate the application of deep learning and machine learning techniques for the analysis of electrohysterogram (EHG) data, which are uterine electrical impulses used to detect preterm deliveries. The TPEHG dataset is used to train a variety of machine learning classifiers, such as Support Vector Machines, Logistic Regression, and Decision Trees, as well as Deep Neural Networks like Convolutional Neural Networks and LSTMs. Additionally, we use plurality voting to build an ensemble of various neural networks and binary classifiers that are trained to classify EHG signals. The ensemble machine learning classifier with five base classifiers produced the best results overall, with an accuracy of 98.99%, sensitivity of 98.3%, and specificity of 97.9% outperforming several state-of-the-art algorithms for preterm birth detection.

The word epilepsy is related to a neurological disease occurred by abnormalities of brain neurons. Timely detection of epilepsy is helpful for patients to decrease the mortality rate. To detect seizures, the Encephalogram (EEG) signals are analyzed based on monitoring the conditions of patients, and seizures can be detected from the EEG signal at appropriate times. The manual detection from the EEG signal requires more time for detecting the seizures and also it needs domain knowledge. The miss detection is eliminated by improving the classification performance in automatic epilepsy detection. Nowadays, deep learning models have not been greatly harnessed in the detection of epileptic seizures due to inappropriate descriptions of time-domain signals and sub-optimal classifier design. The aforementioned issues are combated by the novel Adaptive Hybrid Deep Learning (AHDL) approaches for epilepsy detection using EEG signals. Initially, the required EEG signal is collected from benchmark datasets. The collected signals are subjected to a signal decomposition phase that is accomplished by five levels of decomposition using Dual-Tree Complex Wavelet Transform (DTCWT), where the parameters are tuned by Improved Probability-based Coyote Optimization Algorithm (IP-COA). Further, the decomposed signal is given for feature extraction, where it divides the signal into two phases. In the first phase, the first feature set is obtained by using One-Dimensional Convolutional Neural Network (1DCNN), whereas in the second phase, the proposed model utilizes Auto Encoder (AE) to provide the second feature set. These resultant features are getting fused and the optimal feature selection process is found, where the features are obtained optimally by the IP-COA. Finally, epilepsy detection is accomplished with the aid of proposed AHDL with both Radial Basis-Recurrent Neural Networks (RB-RNN), where the hyperparameters are optimized using IP-COA. Thus, the experimental results illustrate that the suggested model enhances the detection and classification rate.