Automated Tumor Diagnosis with Contrast-Enhanced Ultrasound

项目描述（由申请人提供）：本项目的目标是开发一套计算系统，对超声造影（CEUS）影像进行逐像素分析，量化肿瘤灌注特征，实现高精度、高可靠性的肿瘤病变半自动诊断。在过去的十年中，超声造影通过静脉注射~2¿m充满气体的微泡，引起灌注组织的增强，并允许高度特异性的时间分辨肿瘤血管和灌注成像，显著提高了超声的诊断潜力。然而，超声造影影像的分析需要经验丰富的放射科医生可靠、准确地诊断肿瘤，而观察者之间的可靠性往往是不可靠的

DESCRIPTION (provided by applicant): The goal of this project is to develop a computational system that analyzes contrast-enhanced ultrasound (CEUS) cines on a pixel-by-pixel basis and quantifies tumor perfusion characteristics to semi-automatically diagnose tumor lesions with high accuracy and reliability. Over the past decade, CEUS has significantly improved ultrasound's diagnostic potential by intravenously injecting ~2 ¿m gas filled microbubbles, causing enhancement of perfused tissues and allowing highly specific time-resolved imaging of tumor vasculature and perfusion. However, analysis of CEUS cines requires highly experienced radiologists to reliably and accurately diagnose tumors, and interobserver reliability can often be an issue. A prototype system capable of accurately differentiating benign and malignant tumors has already been developed, tested on an animal tumor model, and published. The proposed research makes use of two CEUS datasets of benign and malignant lesions in the liver and breast from other approved clinical studies at UC San Diego. To achieve clinical success, more sophisticated motion correction algorithms are required because tumors located near the lungs experience motion in multiple directions under breathing. Since image registration techniques can only correct in-plane motion, we will develop a novel through-plane motion filtering technique to focus time-intensity curve analysis on data points collected at the same physical position within the motion cycle. Pixel-by-pixel perfusion kinetic and enhancement measurements will quantitatively and reliably characterize overall tumor behavior and heterogeneity and will form the basis for differentiating tumor types. To develop the computer-aided diagnosis system, we will train a sparse linear discriminant analysis classifier on half of the cines randomly selected from the dataset by minimizing the 10-fold cross-validation error rate. The remaining cines will form the testing set, and the system will be evaluated on its accuracy in correctly differentiating benign versus malignant tumors in the testing set. An automated system to accurately diagnose tumors will allow clinicians to make informed decisions on the best course of action to treat their patients. In addition to allowing optimal patient care, these minimally invasive techniques will be more tolerable to the patient, will reduce unnecessary exposure to DNA-damaging ionizing radiation, and reduce costs to the healthcare system. By eliminating the need for expert radiologists for manually interpretation of CEUS, this medical advancement can help improve access to quality healthcare nationwide, especially in underserved communities where highly experienced radiologists may not be available. Considering the relatively low cost and worldwide acceptance of ultrasound systems, the medical impact of this proposed research extends far beyond the borders of the United States and can potentially improve medical care in less affluent nations.