Fig.1.

Method overview: We provide explanation for the black-box function f(x) interms of concepts, that are radiographic observations mentioned in radiology reports. 1) The intermediate representation Φ₁(x) is used to learn a sparse logistic regression $h_{{\beta }_k}\left(\cdot \right)$ to classify k^th concept. 2) The non-zero coefficients of β_k represents a set of concept units ${\mathcal{V}}_k$ that serves as a mediator in the causal path connecting input x and outcome y. 3) A decision tree function is learned to map concepts to class labels.