Automatic breast cancer risk assessment from digital mammograms

Gopal Raghunath Karemore, Sami Brandt, N Karssemeijer, Mads Nielsen

Publikation: KonferencebidragKonferenceabstrakt til konferenceForskning

Abstract

Purpose:
Textural characteristics of the breast tissue structure on mammogram have been shown to improve breast cancer risk assessment in several large studies. Currently, however, the texture is not used to assess risk in standard clinical procedures or involved in general breast cancer risk assessment tools such as Wolfe Patterns (Wolfe et al 1997), Tabar Patterns (Tabar et al 1982), radiologist’s categorical scorings Breast Imaging Report and Data System® (BIRADS) (ACR 2003), and computer-assisted planimetric measures of area percentage dense tissue (Byng et al 1994). Moreover, these tools rely on a radiologist’s assessment of mammographic appearance that makes it more vulnerable to false negative rate due to observational oversights and varying experience. In addition, these tools are not fully automatic that reduces the workflow efficiency in large screening studies.
In this work, we have investigated a fully automatic and robust risk assessment tool that can take not just the density but also the texture and heterogeneity of the breast tissue into account. By the use of computerized pattern recognition and machine learning techniques, the local texture may be scored for disposition of breast cancer development. We additionally evaluate to which degree the local texture can be recognized to distinguish high risk patients and whether the derived information increases the power of categorical and/or planimetric density scoring.

Materials and Methods
Our cross-sectional case-control study (Otten et al, 2005) includes mammograms (MLO view) of 245 patients diagnosed with breast cancer in the subsequent 2-4 years (123 interval and 122 screen detected cancers) and 250 matched controls. We use the state-of-the-art anatomical breast coordinate system (Brandt et al, submitted) where every pixel location is represented by geodesic distance from nipple and parametric angle, instead of x and y in traditional Cartesian coordinate system within a breast coordinate system, thus locale tissue orientation is compared more accurately (G. Karemore et al, 2010). For every pixel, a collection of multi-scale Gaussian derivative features are extracted in addition to the anatomical locations. Textural information is captured by the Gaussian derivatives and they are used for scoring the mammograms. The scoring includes feature extraction, feature selection, classification and committee learning, as described in detail by Brandt et al (submitted) and Raundahl et al (2008).
Result:
The performance of the proposed imaging marker was compared with various radiologist assisted scoring by area under ROC curve as shown in Table 1 and Fig. 1. Our proposed BC measure showed the maximum area under the ROC curve of 0.64, which is statistically significant and different than other methods (P<0.001, De Long test).
Conclusion:
The proposed fully automatic framework is reproducible and more indicative of breast cancer risk than radiologist’s or computerized density scoring and is also independent from breast density. In future, further validation on other studies should show its general applicability and usefulness in clinical practice.
Clinical trials validation in further studies may improve its general applicability and development.
Keywords: Breast cancer, Imaging marker, Automatic risk scoring.
OriginalsprogEngelsk
Publikationsdato2011
Antal sider1
StatusUdgivet - 2011
BegivenhedEuropean Congress of Radiology - Wien, Østrig
Varighed: 3 mar. 20117 mar. 2011

Konference

KonferenceEuropean Congress of Radiology
Land/OmrådeØstrig
ByWien
Periode03/03/201107/03/2011

Bibliografisk note

Forfatterne ses ikke at have deltaget i konferencen: http://www.myesr.org/cms/website.php?id=/en/past_congresses/ecr_2011.htm

Citationsformater