Texture analysis applied to full-field digital mammography: ability to discriminate between invasive ductal and invasive lobular breast cancer - preliminary results

Purpose

To determine texture features of IDC and invasive lobular carcinoma (ILC) of the breast on full-field digital mammography (FFDM). To evaluate the ability of texture analysis to differentiate between those tumor types.

Materials and methods

Fourteen IDC and nine ILC imaged with FFDM were included in this study. For each lesion the ROI was manually defined covering the lesion and 1 cm normal-appearing breast tissue around the lesion. Texture features derived from the grey-level histogram, the co-occurrence matrix, the run-length matrix, the absolute gradient, the autoregressive model, and the wavelet transform were calculated for the ROIs. Fisher coefficients were calculated to determine which texture features were best suited for distinguishing between IDC and ILC. Based on the combination of those five texture features with the highest Fisher coefficients, lesion classification was performed, using linear discriminant analysis (LDA) and principal component analysis (PCA) classifiers, as well as a k-means clustering algorithm. Classification accuracy was used as the primary outcome measure.

Results

Of the five texture features with the highest Fisher coefficients, the top four were derived from the wavelet transform. Using LDA and PCA, classification accuracies of 82.6% (19 of 23 lesions) and 78.3% (18 of 23 lesions) were achieved, respectively. k-means clustering also yielded a similar classification accuracy of 82.6% (19 of 23 lesions).

Conclusions

Texture features, best suited for discrimination between ILC and IDC, are derived from the wavelet transform. Texture analysis of breast cancer cases imaged with FFDM allows a good degree of accuracy of discrimination between IDC and ILC.