Validation of Automated Breast Density Measurement Program by Comparison with BIRADS Classification and Cumulus®

Abstract

Background: Breast density has been reported as a risk factor of breast cancer. The purpose of this study is to validate the automated breast density measurement program, SIGMAM (Statistics and Information of Gradient Based Mammogram Analysis Method) by comparison with BIRADS classification and manual threshold technique, Cumulus®. Method: SIGMAM performs a pre-processing to extract two features: statistical and edge information. Statistical information is mean and standard deviation of hypothetical fat and dense region in breast area. Edge information is the average of gradient magnitudes from a set of pixels having same intensity. These features are calculated over all pixel intensity range. By combining these two features, the optimal threshold is determined which best divides the fat and dense region. For validation study, 100 cases of Full-Field Digital Mammography (FFDM) taken in Seoul National University Hospital in 2008 were used. Three investigators carried out the validation study. The breast parenchymal densities of 100 images were measured by BIRADS classification, manual mode SIGMAM, automatic mode SIGMAM, and Cumulus®. They were measured by all three investigators. Each investigator repeated the procedure with 1-week interval to check reproducibility and intra-investigator variation. Result: The correlation coefficients of manual mode SIGMAM, automatic mode SIGMAM, and Cumulus® to BIRADS classification were 0.875, 0.802, and 0.904, respectively (Spearman coefficient, p<0.01). The correlation coefficients of manual mode SIGMAM and automatic mode SIGMAM to Cumulus® were 0.934 and 0.877, respectively (Pearson coefficient, p<0.01). The correlation coefficient between manual mode SIGMAM and automatic mode SIGMAM was 0.890 (Pearson coefficient, p<0.01). The correlation coefficients of parenchymal density measured by manual mode SIGMAM among investigators were 0.920, 0.924, and 0.931, respectively (Pearson coefficient, p<0.01). The 1-week interval results by manual mode SIGMAM showed correlation coefficient of 0.975, 0.949, and 0.970 for investigator #1, #2, and #3, respectively (Pearson coefficient, p<0.01). Conclusion: Not only does SIGMAM correlate well with other conventional qualitative and observer-dependent methods, but also the inter- and intra-investigator variation is very low. With its high usefulness and stability, SIGMAM may be utilized in breast density measurement using mammographic images.