FIG 1.

Word-level convolutional neural network for extracting the absence or presence of atypical ductal hyperplasia from a breast pathology report. The free-text report (top left) is turned into a list of embeddings (bottom left). Convolutions, nonlinear projections, and the max over time pooling are then performed to allow the SoftMax function to predict the value probability. The extracted value is finally compared with the annotation in the ground truth. DOB, date of birth.