TY - CONF

T1 - Yet Another Mammography Measure to Evaluate Breast Cancer Risk

AU - Karemore, Gopal Raghunath

AU - Brandt, Sami

AU - Nielsen, Mads

N1 - Conference code: 4

PY - 2009

Y1 - 2009

N2 - Background:Breast density has been shown to improve breast cancer risk assessment in several large studies. Currently, however, the density is not used to assess risk in standard clinical procedures or included 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). All these rely on a radiologist’s assessment of mammographic appearance.  We propose a novel data driven measures taking not just the density but also the texture and its heterogeneity into account. By use of computerized pattern recognition techniques, the local texture may be scored for disposition of breast cancer development.Objective:Investigate to which degree the local texture can be recognised to distinguish high risk patients and whether the derived information increases the power of categorical and/or planimetric density scoring so as to benchmark our new mammography marker.Material 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. The textural information in every pixel is used for scoring the mammogram (Raundahl et al 2008). In every pixel, a collection of multi-scale features are measured at four different scales (1mm, 2mm, 4mm, and 8mm). Specifically the third order horizontal (relative to breast orientation) derivative was used, measuring the anterior-posterior texture component. The center of the breast was defined as the point of largest distance to the boundary. The position relative to the center of the breast was recorded. These statistics are for the individual pixels in a given mammogram compared to statistics of pixels from other mammograms, and it is recorded how many of the 100 most alike pixels found in the other mammograms are from cases and how many are from controls. These counts act as votes for, respectively, high risk and low risk. The sum of votes from all pixels in the mammogram is counted and the percentile of votes for risk is recorded. The offset to 49% is reported. To avoid bias and overtraining issues patients were left out of the statistical analysis when their scores were computed. This was always done leaving a case and a control out simultaneously to treat both classes equal.Results:The categorical score significantly separated cancer and control (p=0.001) and also interval and screen cancers (p=0.04). Similar results were obtained for the area percentage (cancer vs control p<0.0001, screen vs interval cancer p=0.04). Our new measure significantly separated cancer and controls (p<10-8). The power of new this new measure was significantly larger (ROC= 0.65) than BIRADS (ROC=0.58) and area percentage (ROC=0.61).Conclusion:Our new measure is automatic, reproducible and more indicative of risk than radiologist’s or computerized density scoring and is also independent from these. Validation in further studies may prove its general applicability.

AB - Background:Breast density has been shown to improve breast cancer risk assessment in several large studies. Currently, however, the density is not used to assess risk in standard clinical procedures or included 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). All these rely on a radiologist’s assessment of mammographic appearance.  We propose a novel data driven measures taking not just the density but also the texture and its heterogeneity into account. By use of computerized pattern recognition techniques, the local texture may be scored for disposition of breast cancer development.Objective:Investigate to which degree the local texture can be recognised to distinguish high risk patients and whether the derived information increases the power of categorical and/or planimetric density scoring so as to benchmark our new mammography marker.Material 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. The textural information in every pixel is used for scoring the mammogram (Raundahl et al 2008). In every pixel, a collection of multi-scale features are measured at four different scales (1mm, 2mm, 4mm, and 8mm). Specifically the third order horizontal (relative to breast orientation) derivative was used, measuring the anterior-posterior texture component. The center of the breast was defined as the point of largest distance to the boundary. The position relative to the center of the breast was recorded. These statistics are for the individual pixels in a given mammogram compared to statistics of pixels from other mammograms, and it is recorded how many of the 100 most alike pixels found in the other mammograms are from cases and how many are from controls. These counts act as votes for, respectively, high risk and low risk. The sum of votes from all pixels in the mammogram is counted and the percentile of votes for risk is recorded. The offset to 49% is reported. To avoid bias and overtraining issues patients were left out of the statistical analysis when their scores were computed. This was always done leaving a case and a control out simultaneously to treat both classes equal.Results:The categorical score significantly separated cancer and control (p=0.001) and also interval and screen cancers (p=0.04). Similar results were obtained for the area percentage (cancer vs control p<0.0001, screen vs interval cancer p=0.04). Our new measure significantly separated cancer and controls (p<10-8). The power of new this new measure was significantly larger (ROC= 0.65) than BIRADS (ROC=0.58) and area percentage (ROC=0.61).Conclusion:Our new measure is automatic, reproducible and more indicative of risk than radiologist’s or computerized density scoring and is also independent from these. Validation in further studies may prove its general applicability.

M3 - Poster

Y2 - 25 June 2009 through 26 June 2009

ER -