Discrimination analysis of breast calcifications using x-ray dark-field radiography

Background X-ray dark-field radiography could enhance mammography by providing more information on imaged tissue and microcalcifications. The dark field signal is a measure of small angle scattering and can thus provide additional information on the imaged materials. This information can be useful for material distinction of calcifications and the diagnosis of breast cancer by classifying benign and malign association of these calcifications. Methods: For this study, institutional review board approval was obtained. We present the evaluation of images acquired with interferometric grating-based x-ray imaging of 323 microcalcifications (166 malign and 157 benign associated) in freshly dissected breast tissue and compare the results to the information extracted in follow-up pathological evaluation. The number of imaged calcifications is sufficiently higher than in similar previous studies. Fourteen calcification properties were extracted from the digital images and used as predictors in three different models common in discrimination analysis namely a simple threshold model, a naive Bayes model and a linear regression model, which classify the calcifications as associated with a benign or suspicious finding. Three of these fourteen predictors have been newly defined in this work and are independent from the tissue background surrounding the microcalcifications. Using these predictors no background correction is needed, as in previous works in this field. The new predictors are the length of the first and second principle component of the absorption and dark-field data, as well as the angle between the first principle component and the dark-field axis. We called these predictors data length, data width, and data orientation. Results In fourfold cross-validation malignancy of the imaged tissue was predicted. Models that take only classical absorption predictors into account reached a sensitivity of 53.3% at a specificity of 81.1%. For a combination of predictors that also include dark field information, a sensitivity of 63.2% and specificity of 80.8% were obtained. The included dark field information consisted of the newly introduced parameters, data orientation and data width. Conclusions While remaining at a similar specificity, the sensitivity, with which a trained model was able to distinguish malign from benign associated calcifications, was increased by 10% on including dark-field information. This suggests grating-based x-ray imaging as a promising clinical imaging method in the field of mammography.