Robust strain estimation and bi-modal ultrasonic feature extraction methods for improved diagnosis of breast tumors

Abstract

Early detection of breast cancer as well as classi¯cation of breast tumors into benign- malignant and/or BI-RADS ·3, 4, 5 through a noninvasive imaging modality is of prime importance all over the world since the rate of positive ¯ndings at biopsy, the gold standard for diagnosis of breast cancer, is signi¯cantly low. In this dissertation, novel techniques for the detection and classi¯cation of breast tumors are presented employing ultrasound elastography (UE) and conventional B-mode ultrasound (US) based bi-modal imaging with a view to improve the quantitative diagnostic accuracy of breast tumors and thereby reduce the number of unnecessary biopsies. In this work, a new phase-based robust strain imaging algorithm is developed for the detection of breast tumors. The axial strain is calculated by exploiting the direct phase- strain relationship derived from the phase of the zero-lag cross-correlation function between the windowed pre-compression and stretched post-compression analytic RF signals. Unlike conventional phase-based strain estimators, strain is computed in one step using the secant algorithm. To maintain strain continuity, instead of using the instantaneous phase of the interrogative window alone, an average phase function is de¯ned using the phases of the neighboring windows with the assumption that the strain is essentially similar in a close physical proximity of the interrogative window. Comparative results using the simulation and experimental phantom data and in vivo breast data demonstrate that the strain image quality as well as the quantitative image indices of the proposed phase-based elasticity method is better than the other reported techniques. Novel quantitative methods for classi¯cation of 201 breast tumors into benign- malignant employing a bi-modal feature reduction approach in the original as well as in the transform domain are proposed in this dissertation. An original domain bi-modal characterization index is de¯ned from a linear combination of the selected US and UE features along with two new quantitative US features proposed in this work. The weights of linear combination are estimated employing the genetic algorithm (GA). Next, a reduced bi-modal feature based multi-class (i.e., two as well as BI- RADS ·3, 4, 5) classi¯er of breast tumors in the transform domain is proposed. The objective of this part is to construct a minimal set of transform domain features from the de-correlated EMD and DWT coe±cients of an optimally ordered sequence of the original features, employing the wrapper/¯lter approach. Additionally, in order for micro feature-based classi¯cation, two novel spectral domain techniques for estimation of attenuation coe±cient (AC) from the slope of the regression line ¯tted to the 1) modi¯ed average midband ¯t value, and 2) average center frequency shift along depth, Early detection of breast cancer as well as classi¯cation of breast tumors into benign- malignant and/or BI-RADS ·3, 4, 5 through a noninvasive imaging modality is of prime importance all over the world since the rate of positive ¯ndings at biopsy, the gold standard for diagnosis of breast cancer, is signi¯cantly low. In this dissertation, novel techniques for the detection and classi¯cation of breast tumors are presented employing ultrasound elastography (UE) and conventional B-mode ultrasound (US) based bi-modal imaging with a view to improve the quantitative diagnostic accuracy of breast tumors and thereby reduce the number of unnecessary biopsies. In this work, a new phase-based robust strain imaging algorithm is developed for the detection of breast tumors. The axial strain is calculated by exploiting the direct phase- strain relationship derived from the phase of the zero-lag cross-correlation function between the windowed pre-compression and stretched post-compression analytic RF signals. Unlike conventional phase-based strain estimators, strain is computed in one step using the secant algorithm. To maintain strain continuity, instead of using the instantaneous phase of the interrogative window alone, an average phase function is de¯ned using the phases of the neighboring windows with the assumption that the strain is essentially similar in a close physical proximity of the interrogative window. Comparative results using the simulation and experimental phantom data and in vivo breast data demonstrate that the strain image quality as well as the quantitative image indices of the proposed phase-based elasticity method is better than the other reported techniques. Novel quantitative methods for classi¯cation of 201 breast tumors into benign- malignant employing a bi-modal feature reduction approach in the original as well as in the transform domain are proposed in this dissertation. An original domain bi-modal characterization index is de¯ned from a linear combination of the selected US and UE features along with two new quantitative US features proposed in this work. The weights of linear combination are estimated employing the genetic algorithm (GA). Next, a reduced bi-modal feature based multi-class (i.e., two as well as BI- RADS ·3, 4, 5) classi¯er of breast tumors in the transform domain is proposed. The objective of this part is to construct a minimal set of transform domain features from the de-correlated EMD and DWT coe±cients of an optimally ordered sequence of the original features, employing the wrapper/¯lter approach. Additionally, in order for micro feature-based classi¯cation, two novel spectral domain techniques for estimation of attenuation coe±cient (AC) from the slope of the regression line ¯tted to the 1) modi¯ed average midband ¯t value, and 2) average center frequency shift along depth, are proposed. Unlike conventional approaches, this algorithm enforces a controlled continuity in the estimated AC, using the weighted nearest neighbors approach. The highly improved classi¯cation performance of the proposed original domain bi-modal characterization index compared to the conventional bi- or uni-modal indices and even enhanced performance by the transform domain reduced feature vectors in comparison to the original domain ones indicates the potency of both methods in the reduction of unnecessary biopsies. Although the accuracy of the proposed AC estimation algorithm is better than the other reported techniques, the classi¯cation performance of AC alone is not satisfactory yet.