Ultrasound imaging is one of the most effective screening tools for the discovery of tumors. It possesses the advantages of non-invasiveness and irradiation-free, making it a suitable choice in regular screening. By interpreting the ultrasound images, doctors make decisions if the patients requires further examinations or interventions. However, the interpretation of ultrasound images are mostly subjective and highly depend on the experience and judgment of the operator, and the inter-observer variation often results in substantially different decisions. Due to these reasons, how to objectively and quantitatively describe a tumor in ultrasound imaging becomes a pressing issue facing the medical staffs.
To effectively quantify the sonographic features, an objective accurate tumor location and boundary have to be first determined. Therefore, computer-aided systems in our research are respectively proposed to automatically segment boundary of tumor and detect tumor location on ultrasound images.
1.1 Boundary segmentation for ultrasonic thyroid nodules
1.1.1 Background
Thyroid cancer has increased significantly over the past four decades in United States [1]. The advance in diagnostic tools allows the detection of small thyroid nodules [1]-[4].
Although early treatment is the key to the cure of cancer and further reduces the mortality rate [3], it may also lead to over-diagnosis with unnecessary biopsies and treatments [1]-[4]. Therefore, an appropriate screening tool with effective diagnosis ability has become an increasingly important issue [5]-[7].
Ultrasound imaging is one of the most effective non-invasive screening tools for the detection of thyroid nodules [6]-[14]. Beside conventional B-mode sonography,
elastography which reflects tissue stiffness plays a promising role in the characterization of thyroid nodules [17]. Based on impressions of observing the ultrasound images, clinicians make suggestions for patients to be subject to periodic follow-up or further cytological tests [8]. However, the acquisition and observation of ultrasound images are mostly subjective and highly depend on the experience and judgment of the operator, and the inter-observer variation often results in significantly different decisions [6], [11]-[13].
According to aforementioned reasons, objective quantification of the sonographic thyroid nodule features has become a pressing issue facing the medical staffs [6]. To effectively quantify the sonographic features, an objective accurate nodule boundary has to be first determined. Therefore, computer-aided system was proposed to automatically detect and identify the boundary of thyroid nodules on ultrasound images [9]-[14].
1.1.2 Objective
For computer-aided systems, speckle is one of the most challenging problems [9]-[14], [15], [16]. To reduce the impact of speckle, previous publications used image preprocessing methods or altered the detailed information provided during the execution steps [9]-[11], [14]. However, speckle also carries information about tissue characteristics and may be exploited in several applications. Approaches in attempt to reduce speckle may result in the loss of such information within original ultrasound images, and may even affect the capturing of nodule features and the correctness of the computer-aided system operation.
To the best of our knowledge, this is the first work to apply Variance-Reduction (V-R) statistics-based approach to detect nodule boundaries on thyroid ultrasonography. In this study, we propose a novel and semi-automatic method without image preprocessing.
1.2 Breast tumor detection in 3D ultrasound imaging 1.2.1 Background
Breast cancer has the highest incidence and causes the second leading mortality of cancer among women in 2017 in the United States [18]. Since early treatment is the key to the cure of breast cancer, effective diagnosis system and the adjunctive detection system have become an important issue to further reduce the mortality rate [19].
Mammography [20] and breast ultrasound [21] are two of the most commonly used screening tools for the diagnosis and detection of breast tumor. Each of them has its own advantage and disadvantage in clinical examination [22]. Mammography is adopted as a major screening tool for its high sensitivity in detecting micro-calcification and can be used to further detect early changes in abnormal tissues, but it also suffers from high false positive rate especially in women with dense breast tissue [23]-[26]. Breast ultrasound is the most important alternative to mammography [23]-[25], with the advantages of radiation-free and less pain, but ultrasound imaging highly depends on the medical staff’s experience and judgment, resulting in inter-observer variations that often lead to significantly different clinical decisions, and the results are poorly reproducible [26].
Based on these problems, automated breast volume scanner (ABVS) as an innovative high-end ultrasound scanner has been developed by combining the advantages of mammography and breast ultrasound to automatically scan the entire breast [9]-[11].
Additionally, ABVS allows the scan results to have less operator dependency and higher reproducibility, and provides coronal images to further assist the planning of surgical intervention [26], [27]. However, the reviewing process to discover suspicious abnormalities from hundreds of image slices produced by ABVS is often time-consuming [29], [30]. Besides, tumor which is tiny or isoechogenic to the surrounding normal tissue
may be missed by reviewer [30]. Therefore, computer-aided detection (CADe) system has been proposed to accelerate the reviewing process and to reduce the missing errors [33]-[39]. As defined by the U.S. FDA, “CADe devices are computerized systems that incorporate pattern recognition and data analysis capabilities and are intended to identify, mark, highlight, or in any other manner direct attention to portions of an image, or aspects of radiology device data, that may reveal abnormalities during interpretation of patient radiology images or patient radiology device data by the intended user” [40]. CADe system is needed to solve a practical clinical problem of ABVS where thousands of images were acquired by automatically scanning the patient breasts and will be time-consuming and tiresome for physicians to review the slice one by one to discover any suspicious lesions.
1.2.2 Objective
The aim of this study is to develop an intuitive, simple, efficient CADe solution to quickly screen all the images generated by the automated three-dimensional breast ultrasound (ABUS), in particular Siemens ACUSON2000 ABVS in this study, to highlight all possible lesions for further review and diagnosis by physicians.