Abstract: In a method of generating a neural network used to detect a feature of medical significance from a body image data input, test data images (110) are divided into patches (120). Each patch is labelled as either corresponding to the feature (124) or not corresponding to the feature (122). One trained fully connected layer (136) in a pretrained general purpose convolutional neural network (130) is replaced with a new fully connected layer. The pretrained convolutional neural network (130) is retrained with the set of labelled patches (126) to generate a feature-specific convolutional neural network that includes at least one feature-specific fully connected layer (152) that maps the body image data to the feature of medical significance when the feature of medical significance is present in the body image data input.