Antenna design for breast cancer detection and machine learning approach for birth weight prediction

Abstract

With the advancement of technology in the biomedical field, new diagnostic and treatment methods and new devices are being developed day by day. However, although this situation seems mostly advantageous, the development of technology in some areas poses some difficulties for both patients and doctors in terms of diagnosis and treatment. For example, electromagnetic radiation used for diagnostic purposes can be harmful to patients. In addition, the precision and accuracy of the results of the techniques used also contain a margin of error, and it becomes important for doctors to consider these margins of error in the decision-making process. Based on the briefly mentioned problems, alternative methods are proposed for two different fields in this thesis. In the first study, an alternative method different from standard methods for breast cancer diagnosis will be proposed, and in the second study, machine learning approaches that can determine the baby's birth weight with high accuracy will be presented. Breast cancer remains a major global health problem and requires continuous improvements in diagnostic and control methods to achieve better patient outcomes during treatment and early detection of the disease. As breast cancer is one of the most common and dangerous diseases among women worldwide, it is therefore critical to diagnose it quickly. Considering that breast cancer is the second-leading cause of cancer-related mortality in women, the need for efficient and non-invasive diagnostic methods has become greater. The negative consequences of conventional approaches in terms of their operating principles or application methodologies give rise to this demand. In response to the limitations inherent in traditional diagnostic techniques, microwave imaging methods have been developed for effective diagnosis of breast cancer. The feasibility and efficacy of using microstrip patch antennas for breast cancer detection are especially examined in the first section of this thesis, which explores an alternative medical method. These antennas can be considered an important development in the medical industry as they are able to detect small electromagnetic oscillations that are indicative of early-stage cancer. This study introduces the design and simulation of a rectangular microstrip patch antenna on an FR-4 substrate operating at 2.45 GHz in the ISM band for breast cancer detection. Utilizing the Computer Simulation Technology (CST) software, both the proposed antenna and a five-layer breast phantom, with and without a 5 mm-radius tumor, were comprehensively designed. A breast phantom modeled as a hemisphere and an embedded tumor modeled as a sphere with different dielectric characteristics were successfully simulated. The antenna's performance was evaluated at varying distances from the phantom, revealing alterations in parameters such as electric field, return loss, voltage standing wave ratio, efficiency, specific absorption rate, etc., in the presence of a tumor. The simulation results at different antenna locations show discernible differences in values with and without tumors, indicating that a tumor significantly influences power reflection back to the antenna. The VSWR of the antenna, lower than 2, aligns with acceptable VSWR limits. Furthermore, the proposed antenna demonstrates increased electric field strength in the presence of a tumor. In addition, simulation outcomes in free space and with a 3-D breast phantom indicated that the antenna, positioned 20 mm from the breast phantom, is more efficient in tumor identification compared to the one located at 40 mm. Given its tumor detection capability and satisfactory SAR values, the proposed antenna emerges as a promising tool in biomedical applications. Future studies will explore alternative antenna geometries and techniques to enhance performance and increase tumor detection sensitivity. Birth weight is a critical indicator of both pregnancy progress and infant development, exerting a substantial influence on short- and long-term health conditions in newborns. In other words, fetal weight emerges as a pivotal indicator of short- and long-term health problems in newborns, both in developed and developing countries. Understanding the contributing factors to low birth weight (LBW) and high birth weight (HBW) can inform the implementation of optimal interventions for the population's health. In the second study, we present our research on the prediction of birth weight classification through the application of various machine learning algorithms. For this investigation, 913 medical observation units, each characterized by 19 features encompassing actual birth weight information and ultrasound measurements, were employed. In the study, a number of data preprocessing steps were performed on the data set before the data set was directly used to train the classifier models. To address the issue of imbalanced data across classes, we implemented the synthetic minority oversampling technique (SMOTE). Additionally, feature scaling was applied to standardize numerical attributes within a particular range in the dataset, as there are different physiological variables with different units and orders of magnitude. In this work, nine different machine learning classifier models are used. They are decision tree, discriminant analysis, naive bayes, support vector machine, k-nearest neighbor, kernel approximation, ensemble classifier, artificial neural network, and logistic regression. The hyperparameters of each model were kept at default values, and no hyperparameter tuning was made. To evaluate the performance of nine distinct supervised machine learning algorithms, we compared birth weight classification models with and without feature selection, utilizing numerous evaluation metrics. These different metrics are accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, and area under the receiver operating curve. Referring to the Pearson correlation coefficient technique applied to the data set, abdominal circumference, head circumference, biparietal diameter, femur length, and hemoglobin levels at the 0th and 6th hours are highly correlated with birth weight. The results of our analysis highlight that the subspace kNN-based ensemble classifier outperforms other machine learning models, achieving the best macro-average accuracy of 99.87% without feature selection and 99.75% with feature selection. Additionally, we observed that the bilayered neural network exhibits similar performance to the kNN-based model, with the best macro-average accuracy of 99.62%, irrespective of feature selection. Furthermore, principal component analysis (PCA) was applied to the data set as an unsupervised method for birth weight classification. The outcome clearly demonstrates the successful classification of most data points by PCA. The findings of this study underscore the potency of machine learning as a robust and non-invasive method for accurately predicting the birth weight classification of infants. In light of these factors, a health program could be devised to prevent the occurrence of LBW and HBW since recognition of LBW or HBW in a newborn may signal potential problems that could manifest immediately after birth or later in life. At the end of the thesis, performance improvement methods have been proposed based on the two studies we conducted, and we hope that the results of our research will shed light on future studies.