Ultra wideband technique for breast cancer detection using multi-layer feed-forward neural networks

Breast cancer is one of the main causes of women‘s death. Early detection of tumors increases the chances of overcoming this disease. There are several diagnostic methods for detecting tumors, each of which has its own limitations. Recently, Ultra Wideband (UWB) imaging has gained wide acceptance for several good features such as its specificity and lack of ionizing radiation. The confocal method has been the dominant technique in this area based on homogeneous breast tissues and prior knowledge of tissue permittivity. Hence it is impractical and difficult to be implemented clinically. This thesis has focused on development of a complete non-confocal system for breast tumor detection using Neural Network (NN)-based Ultra Wideband (UWB) imaging considering both homogeneous and heterogeneous tissues. The work has been done in two phases: i) Simulation and ii) Experiment. At the simulation stage, a feed-forward NN model was developed to identify the existence, size, and location of tumors in a breast model. Spherical tumors were created and placed at arbitrary locations in a hemispherical breast model using the Computer Simulation Technology (CST) software as an Electromagnetic (EM) simulator. The UWB signals were transmitted and received through breast phantoms. The transmitter and receiver were rotated 360° to detect tumor existence, size, and location in a two-dimensional breast slice using the best-complement rule. A modified Principle Feature Analysis (PFA) method was implemented to reduce the feature vector size and extract the most informative features. We have found that the most informative features occur at the maxima and minima of the signals. The extracted features from the received UWB signals were fed into the NN model to train, validate, and test it first and then to detect the presence, size, and location of possible breast tumors. After simulation proof, a system was developed for experimental tumor detection. The system consisted of commercial UWB transceivers, a developed NN model, and breast phantoms for homogenous and heterogeneous tissues. The breast phantoms and tumor were constructed using available low cost materials and their mixtures with minimal effort. The materials and their mixtures were chosen according to the ratio of the dielectric properties of the breast tissues. A Discrete Cosine Transform (DCT) of the received signals was used to construct the feature vector to train the NN model. Finally, the system was trained to distinguish between malignant and benign tumors. Tumors as small as 0.1 mm and 0.5 mm (diameter) have been successfully detected through simulation and experimental investigation respectively. The tumor existence, size, and location detection rate are about (i) 100%, 93%, and 93.3% and (ii) 100%, 95.8%, and 94.3% through simulation and experimental system respectively. Possible differentiation between malignant and benign tumor was also achieved. The method utilizes the power of neural networks and demonstrates a new direction in this field. This gives assurance of breast tumor detection and the practical usefulness of the developed system in the near future.

Preview PDF FK 2011 71R.pdf Download (956kB) | Preview

Actions (login required)

View Item