Approximate Explanations for Classification of Histopathology Patches

An approximation method for faster generation of explanations in medical imaging classifications is presented. Previous results in literature show that generating detailed explanations with LIME, especially when fine tuning parameters, is very computationally and time demanding. This is true both for manual and automatic parameter tuning. The alternative here presented can decrease computation times by several orders of magnitude, while still identifying the most relevant regions in images. The approximated explanations are compared to previous results in literature and medical expert segmentations for a dataset of histopathology images used in a binary classification task. The classifications of a convolutional neural network trained on this dataset are explained by means of heatmap visualizations. The results show that it seems to be possible to achieve much faster computation times by trading off finer detail in the explanations. This could give more options for users of artificial intelligence black box systems in the context of medical imaging tasks, in regards to generating insight or auditing decision systems.