Computer-aided image quality assessment in automated 3D breast ultrasound images
|Other Titles:||Computer-gestütze Bildqualitätsanalyse für automatische 3D Brust-Ultraschallbilder||Authors:||Schwaab, Julia||Supervisor:||Günther, Matthias||1. Expert:||Günther, Matthias||2. Expert:||Marti, Robert||Abstract:||
Automated 3D breast ultrasound (ABUS) is a valuable, non-ionising adjunct to X-ray mammography for breast cancer screening and diagnosis for women with dense breasts. High image quality is an important prerequisite for diagnosis and has to be guaranteed at the time of acquisition. The high throughput of images in a screening scenario demands for automated solutions. In this work, an automated image quality assessment system rating ABUS scans at the time of acquisition was designed and implemented. Quality assessment of present diagnostic ultrasound images has rarely been performed demanding thorough analysis of potential image quality aspects in ABUS. Therefore, a reader study was initiated, making two clinicians rate the quality of clinical ABUS images. The frequency of specific quality aspects was evaluated revealing that incorrect positioning and insufficiently applied contact fluid caused the most relevant image quality issues. The relative position of the nipple in the image, the acoustic shadow caused by the nipple as well as the shape of the breast contour reflect patient positioning and ultrasound transducer handling. Morphological and histogram-based features utilized for machine learning to reproduce the manual classification as provided by the clinicians. At 97 % specificity, the automatic classification achieved sensitivities of 59 %, 45 %, and 46 % for the three aforementioned aspects, respectively. The nipple is an important landmark in breast imaging, which is generally---but not always correctly---pinpointed by the technicians. An existing nipple detection algorithm was extended by probabilistic atlases and exploited for automatic detection of incorrectly annotated nipple marks. The nipple detection rate was increased from 82 % to 85 % and the classification achieved 90 % sensitivity at 89 % specificity. A lack of contact fluid between transducer and skin can induce reverberation patterns and acoustic shadows, which can possibly obscure lesions. Parameter maps were computed in order to localize these artefact regions and yielded a detection rate of 83 % at 2.6 false positives per image. Parts of the presented work were integrated to clinical workflow making up a novel image quality assessment system that supported technicians in their daily routine by detecting images of insufficient quality and indicating potential improvements for a repeated scan while the patient was still in the examination room. First evaluations showed that the proposed method sensitises technicians for the radiologists' demands on diagnostically valuable images.
|Keywords:||Breast ultrasound, ABUS, Breast cancer screening, Image processing, Machine learning, Random Forest||Issue Date:||22-Jan-2016||URN:||urn:nbn:de:gbv:46-00104967-13||Institution:||Universität Bremen||Faculty:||FB1 Physik/Elektrotechnik|
|Appears in Collections:||Dissertationen|
checked on Sep 19, 2020
Items in Media are protected by copyright, with all rights reserved, unless otherwise indicated.