Explanations for deep neural network predictions in terms of domain-related\nconcepts can be valuable in medical applications, where justifications are\nimportant for confidence in the decision-making. In this work, we propose a\nmethodology to exploit continuous concept measures as Regression Concept\nVectors (RCVs) in the activation space of a layer. The directional derivative\nof the decision function along the RCVs represents the network sensitivity to\nincreasing values of a given concept measure. When applied to breast cancer\ngrading, nuclei texture emerges as a relevant concept in the detection of tumor\ntissue in breast lymph node samples. We evaluate score robustness and\nconsistency by statistical analysis.\n