A plurality of mammograms 10 is received, a first analysis 30 identifies a malignancy classification for each of the mammograms； a malignancy output value 30Y is determined for each of the mammograms dependent on the first analysis； an average malignancy output value is determining by averaging the malignancy output values for the plurality of mammograms； the average malignancy output value is thresholded to generate a output binary malignancy value 60Y； a second analysis 40 determines a plurality of localisation parameters 40X for each mammogram； output localisation data is generated 70 for the plurality of mammograms in dependence on the output binary malignancy value. The invention seeks to provide an automated malignancy determination in parallel to a human operator to reduce the need for two human operators in a mammography analysis workflow. The plurality of mammograms may be pre-processed using trained neural networks to improve classification. The second analysis may be conducted using a trained regional convolutional neural network (RCNN)； the RCNN may comprise sub-divisional networks which provide: a bounding box generating model and/or a segmentation model and/or a malignancy classification type model.
