Date of Award
Doctor of Philosophy (PhD)
According to National Breast Cancer Society, one in every eight women in United States is diagnosed with breast cancer in her lifetime. American Cancer Society recommends a semi-annual breast-cancer screening for every woman which can be heavily facilitated by the availability of low-cost, non-invasive diagnostic method with good sensitivity and penetration depth. Ultrasound (US) guided Diffuse Optical Tomography (US-guided DOT) has been explored as a breast-cancer diagnostic and screening tool over the past two decades. It has demonstrated a great potential for breast-cancer diagnosis, treatment monitoring and chemotherapy-response prediction. In this imaging method, optical measurements of four different wavelengths are used to reconstruct unknown optical absorption maps which are then used to calculate the hemoglobin concentration of the US-visible lesion. This dissertation focuses on algorithm development for robust data processing, imaging reconstruction and optimal breast cancer diagnostic strategy development in DOT. The inverse problem in DOT is ill-posed, ill-conditioned, and underdetermined. This makes the task of image reconstruction challenging, and thus regularization-based method need to be employed. In this dissertation, a simple two-step reconstruction method that can produce accurate image estimates in DOT is proposed and investigated. In the first step, a truncated Moore-Penrose Pseudoinverse solution is computed to obtain a preliminary estimate of the image. This estimate can be reliably determined from the measured data; subsequently, this preliminary estimate is incorporated into the design of a penalized least squares estimator that is employed to compute the final image estimate. Using physical phantoms, the proposed method was demonstrated to yield more accurate reconstruction compared to other conventional reconstruction methods. The method was also evaluated with clinical data that included 10 benign and 10 malignant cases. The capability of reconstructing high contrast malignant lesions improved by the use of the proposed method.Reconstructed absorption maps are prone to image artifacts from outliers in measurement data from tissue heterogeneity, bad coupling between tissue and light guides, and motion by patient or operator. In this dissertation, a new automated iterative perturbation correction algorithm is proposed to reduce image artifacts based on the structural similarity index (SSIM)) of absorption maps of four optical wavelengths. The SSIM was calculated for each wavelength to assess its similarity with other wavelengths. Absorption map was iteratively reconstructed and projected back into measurement space to quantify projection error. Outlier measurements with highest projection errors were iteratively removed until all wavelength images were structurally similar with SSIM values greater than a threshold. Clinical data demonstrated statistically significant improvement in image artifact reduction.US guidance with DOT helps to reduce false positive rate and hence reduce number of unnecessary biopsies. However, DOT data processing and image reconstruction speed remains slow compared to real-time US. Real-time or near real time diagnosis with DOT is an important step toward the clinical translation of the US-guided DOT. In this dissertation, to address this important need, we present a two-stage diagnostic strategy that is computationally efficient and accurate. In the first stage, benign lesions are identified in near real-time by use of a random forest classifier acting on the DOT measurements and radiologistsՠUS diagnostic scores. The lesions that cannot be reliably classified by the random forest classifier will be passed on to the image reconstruction stage. Functional information from the reconstructed hemoglobin concentrations is used by a Support Vector Machine (SVM) classifier for diagnosis in the second stage. This two-step classification approach that combines both perturbation data and functional features results in improved classification, as quantified using the receiver operating characteristic (ROC) curve. Using this two-step approach, area under the ROC curve (AUC) is 0.937 屠0.009 with sensitivity of 91.4% and specificity of 85.7%. While using functional features and US score, AUC is 0.892 屠0.027 with sensitivity of 90.2% and specificity of 74.5%. The specificity increased by more than 10% due to the implementation of the random forest classifier.
Abhinav Jha, Adam Bauer, Ulugbek Kamilov, Mark Anastasio,