Abstract:
Quantitative ultrasound (QUS) based parameters, estimated from backscattered radio-frequency (RF) data, allows the parametrization of tissue micro-structures. These pa-rameters, which include mean scatterer spacing (MSS) and effective scatterer diameter (ESD), often reveal more information about the interrogated tissue than conventional B-mode imaging, as well as being less subjective to operator settings and interpreter variability than conventional ultrasound. MSS and ESD are important QUS micro-parameters for detecting pathological changes in breast tissue. In this thesis, two novel techniques are proposed for estimation of MSS and ESD from breast tissues. Both these techniques rely on the separation of the coherent and diffuse component of backscattered data using ensemble empirical mode decomposition (EEMD) of the data into their intrinsic mode functions (IMFs). An automatic IMF selection scheme is employed, which utilizes a non-parametric Kolmogorov-Smirnov (K-S) test to automatically select the IMFs responsible for coherent scattering in case of MSS estimation, and diffuse scattering in case of ESD estimation. Before EEMD can be performed, filtering and deconvolution of the backscattered data is carried out to reduce the impacts of diffraction and the system point spread function (PSF). The MSS is estimated from the spacing between the peaks of the power spectrum estimated from the coherent component of RF data. The power spectrum is estimated using an autoregressive (AR) model, whose order is chosen by minimization of a novel mean absolute percentage error (MAPE) criterion. The ESD is estimated from the diffuse component of RF data utilizing a theoretical tissue scattering model in the frequency domain. MSS estimation is carried out on simulation RF data generated by FIELD II and in vivo breast tissues while ESD estimation is carried out on experimental tissue-mimicking phantoms (TMPs) and in vivo tissues. The average MSS for normal tissues, inflammatory tissues, fibroadenoma, and malignant tissues are found to be 0.689 (±0.032) mm, 0.729 (±0.040) mm, 0.750 (±0.035) mm, and 0.793 (±0.040) mm, respectively. The corresponding average ESD values are 75.12 (±4.01) µm, 75.72 (±4.09)
xi
ABSTRACT xii
(±9.55) µm, and 123.05 (±8.85) µm, respectively. The estimated average MSS and
ESD values correspond well to those previously reported in literature. When MSS and
ESD are combined with 27 previously reported QUS bi-modal (ultrasound B-mode
(UB) and ultrasound elastography (UE)) macro-parameters, to form a unique hybrid
micro-macro feature set, consisting of 29 parameters, for binary (benign-malignant)
classification of breast lesions, we obtain sensitivity, specificity, and accuracy, values as
high as 98.21%, 98.06%, and 98.11%, respectively, using machine learning algorithms.
This highlights the potential of this hybrid feature set as a computer-aided diagnosis
(CAD) tool for breast lesion classification.
