UC San Diego

In the fight against cancer, early detection and accurate diagnosis of tumors are critical. Advances over the past decade with contrast-enhanced ultrasound (CEUS) have enabled real time imaging of tumor vasculature, improving ultrasound's diagnostic potential. Additionally, CEUS is cheaper and safer than other medical imaging modalities, but analysis of CEUS requires highly experienced radiologists to accurately and reliably diagnose tumors. To overcome this issue, a computer-aided diagnosis (CAD) system for CEUS was developed to differentiate benign and malignant tumors. A prototype CAD system was developed using data from a murine breast tumor model. CEUS cines of microbubble bolus injections were acquired continuously to capture microbubble wash-in and wash-out. The time- intensity curves were analyzed on a pixel-by-pixel basis to measure kinetic parameters throughout the tumor. Linear discriminant analysis was used to differentiate benign and malignant tumors, achieving 100% cross-validation accuracy. While traditional region-of-interest analyses could measure general tumor perfusion, pixel-by-pixel analysis made possible the detection of blood flow heterogeneity, which was shown to improve tumor differentiation. Although results in animal tumor models were promising, human clinical CEUS was challenging for CAD. Clinical CEUS acquired during free breathing suffer from in-plane and out-of-plane motion, reducing accuracy of quantitative measurements. To reduce motion artifacts, 2-tier in-plane motion correction (IPMC) and out-of-plane motion filtering (OPMF) algorithms were designed and tested on CEUS of focal liver lesions (FLLs). The 2-tier IPMC strategy provided stable motion correction and OPMF reduced apparent motion throughout the cine. These algorithms significantly improved visual stability and quantitative analysis of tumor perfusion. Finally, a CAD system was developed on clinically acquired CEUS bolus injection cines of FLLs. The previously formulated quantitative techniques were combined with newly developed algorithms to detect enhancement and flow patterns known to differentiate FLLs. Support vector machines and artificial neural networks were employed to classify lesions as benign or malignant, achieving 84.6% accuracy in the untrained testing set. The methods developed here have great potential to improve cancer care globally by aiding physicians to differentiate the disease and determine the optimal treatment plan, thus allowing more patients to receive the care that they need