Processing and imaging techniques for microwave-based head imaging

Abstract

Microwave imaging systems for medical applications have been widely investigated in recent years due to their potential to provide portable diagnostic tools that are safe, low cost, and nonionizing. Among many medical applications, brain stroke detection and classification using microwave techniques has been attracting an increasing interest due to the need for a portable onsite, real-time stroke diagnosis that can be used by paramedics. A complete microwave imaging system includes hardware components and software algorithms. The processing and imaging techniques, which are the topic of this thesis, use the collected microwave signals via the antenna array to generate the images of the brain. Numerous microwave imaging techniques applied on bio-medical applications have been researched during recent years. Those proposed techniques exhibited great potential, however, they suffer from several serious drawbacks that need to be solved. This thesis aims to solve four main problems (including a large number of antenna elements, a large number of frequency samples, sensitivity to initial guess of the effective dielectric properties of the image domain and sparsity of the imaged domain), in current microwave imaging techniques and in doing so makes four main research contributions. The first contribution is the development of a novel algorithm based on compressive sensing (CS). The main target is to develop CS-based imaging algorithm to reduce the number of antennas used in the array. A CS model is constructed based on confocal imaging algorithm, and a convex optimization problem is solved in order to reconstruct the reflection coefficients of the imaging domain. The proposed algorithm is successfully tested on a head model. Followed by that, another CS-based algorithm is proposed to reduce the number of stepped frequencies used in the microwave transceiver system. A CS model is constructed based on the sparse time domain signal received by the antenna array. The algorithm is tested using a developed microwave head imaging system. The results indicate that by using 25% of the original stepped frequencies, the image can be ideally recovered by using the proposed algorithm. The second contribution is the development of an optimization based confocal imaging algorithm. In all of the traditional confocal imaging algorithms, the effective dielectric constant of the imaging area has to be initially estimated. Since the generated image is sensitive to the effective dielectric constant, a small error in the estimation will cause large distortion in the image. The proposed algorithm proposes a novel concept in which the effective dielectric constant is considered a variable that depends on the signal’s entry point in the imaged object (the head). Based on this concept, multiple effective dielectric constants are optimized with the aim to achieve the most focused (best) image. This optimization is implemented by using particle swarm optimization method. The proposed method is compared with traditional confocal imaging algorithms. The results indicate that the proposed algorithm can achieve much better images with lower cluster effects and insensitive to the initial values of the effective dielectric constants. The third contribution is the development of a CS based tomography method. A critical problem, which is the sparsity of the imaging domain, is firstly investigated. Since the electrical properties of human head is non-sparse, the wavelet transform is used to transform the non-sparse profile into a sparse wavelet domain. After that, a CS based algorithm named block sparse Bayesian learning (BSBL) is used to recover the electrical properties of the head by using less number of antennas. The proposed algorithm is compared with other traditional tomography methods, and the results indicate that since less number of antennas is used in the system, the images generated by using traditional methods are largely distorted because less information is obtained from the received signal. However, by using the proposed algorithm, the electrical properties of the target area can be ideally recovered by using only 4 antennas with satisfactory results whereas traditional methods require at least 32 antennas. The last contribution is the development of a framework for brain stroke classification. Two databases are firstly constructed by using two numerical head phantoms. The first database is used to train the classifier whereas the second database is used to evaluate the performance of the built classifier. The databases are composed with brain images generated by using Born iterative method. The validity of the framework is verified with various strokes (haemorrhagic or ischaemic) with different sizes and positions. Two machine learning based techniques named K-means clustering and support vector machine are used to build the classifier. The constructed classifier is tested by using the second database, and the results indicate that when different noise levels are considered, the proposed framework can achieve 88% accuracy rate. The receiver operator characteristic curve is also used to test the framework and the results indicate that the framework can successfully localize the stroke and achieve 91% sensitivity and 87% specificity.