Search results

(1 - 1 of 1)

Title

CASE-ADAPTIVE PROCESSING FOR IMPROVING ACCURACY IN COMPUTER-AIDED DIAGNOSIS OF BREAST CANCER

Creator

Sainz De Cea, Maria Victoria

Date

2018, 2018-05

Description

Breast cancer is the most commonly diagnosed cancer among women (apart from skin cancer) in the US. If detected early, the five-year survival...
Show moreBreast cancer is the most commonly diagnosed cancer among women (apart from skin cancer) in the US. If detected early, the five-year survival rate is 99%. Because of this, early detection of breast cancer has been an extensively studied topic over the years, and screening mammography is the gold standard for this purpose. Microcalcifications (MCs) are tiny calcium deposits that appear as bright spots in mammogram images, and they can be an early sign of breast cancer in asymptomatic women. Computer aided diagnosis (CAD) tools can be used to assist radiologists in detecting MCs and classifying them as benign or malignant. CAD of breast cancer is often hampered by the presence of false positives (FP) among the detected MCs when a reasonable sensitivity level is achieved. The FPs can be caused by MC-like noise, linear structures, etc. Due to the wide range of factors causing FPs, there is a great inter-patient variability, which can degrade the performance of CAD systems. In this work, we aim to reduce the inter-patient variability of CAD systems in order to improve the performance in both MC detection (Computer aided detection or CADe) and classification of MC clusters (Computer aided diagnosis or CADx). The first part of this thesis focuses on MC detection. We first develop a framework for estimating the accuracy in detection of individual MCs within a lesion region. This framework is general and can be applied to any MC detector. The number of FP detections can vary greatly from patient to patient, so having this knowledge will be useful to make decisions in both CADe and CADx systems. Secondly, we present a case-adaptive method for CADe based on Bayes’ risks, where a distribution is fit to the FPs from a mammogram under consideration, based on which the optimal detection threshold is determined for each patient. Finally we present an outlier approach for detection of individual MCs in a lesion region. This approach is based on the fact that individual MCs are usually different from the FPs (brighter, larger in extent), so they can be detected as statistical outliers. The outlier detection is done in a case-by-case basis, which can yield not only a reduction in the number of FPs but also an increase on the uniformity of the detection accuracy among different cases. The second part of the thesis is focused on CADx. We apply the methods developed in the first part to improve the uniformity and performance in the classification of detected lesions as benign or malignant. For this purpose we first present a quality factor approach for adjusting the contribution of the detected individual MCs to the final feature set. Those detections with a higher quality factor can have more impact in the final features, therefore mitigating the effect of the FP detections. Finally, we use the estimated detection accuracy to determine the optimal detection operating threshold. This is shown to boost the CADx performance.
Ph.D. in Electrical Engineering, May 2018
Show less