Diagnostic value of apparent diffusion coefficient lesion texture biomarkers in breast MRI

Abstract

Quantitative assessment of breast intra-lesion heterogeneity in terms of contrast agent free Magnetic Resonance Imaging (MRI) approaches hold potential in breast cancer diagnosis. This study focuses on an Apparent Diffusion Coefficient (ADC) based approach, investigating the diagnostic role of 1st and 2nd order ADC statistics features, in differentiating benign from malignant breast lesion status. A total of 67 patients with 78 histologically verified breast lesions (40 benign and 38 malignant) was analyzed. ADC maps were generated for a slice representative of lesion largest diameter, considering intra Diffusion Weighted Imaging (DWI) sequence non rigid registration scheme. Lesion segments were defined by semi-automated Fuzzy C-Means (FCM) segmentation on high b-value diffusion images and propagated on ADC maps. 27 (11 1st order statistics and 16 2nd order statistics (texture) features were derived. To avoid overfitting a stepwise feature selection method was employed, while the discriminating ability of features was evaluated with univariate and multivariate Logistic Regression classification. The classification performance of the diagnostic model was evaluated by means of the Area Under Receiver Operating Characteristic curve (Az index). A combination of two features, one from 1st order statistics (25th Percentile) and one from 2nd order statistics, (texture Entropy), achieved high classification performance (Az = 0.965 ± 0.024), suggesting both the diagnostic significance of 1st order statistics and texture biomarkers of the ADC map representation.