Figure 2.

Diagrammatic illustration of steps involved in the robustness assessment, classification evaluation method. Texture features are first clustered and assessed in terms of robustness using only feature values and vendor information, remaining blinded to risk classification. The union of features identified by clustering features from M1 (machine one) and M2 (machine two) is the set considered to be robust and nonredundant. The most robust and nonredundant features are identified, and only these features are used as feature candidates in classification evaluation. Solid and dashed arrows show two different data pathways followed to evaluate the generalization of classification of the heterogeneous image datasets. The full analysis was repeated twice; once with the GE unit as M1 and the Hologic unit as M2, and then again but with the GE unit as M2 and the Hologic unit as M1.