Explaining COVID-19 and Thoracic Pathology Model Predictions by Identifying Informative Input Features

Neural networks have demonstrated remarkable performance in classification\nand regression tasks on chest X-rays. In order to establish trust in the\nclinical routine, the networks' prediction mechanism needs to be interpretable.\nOne principal approach to interpretation is feature attribution. Feature\nattribution methods identify the importance of input features for the output\nprediction. Building on Information Bottleneck Attribution (IBA) method, for\neach prediction we identify the chest X-ray regions that have high mutual\ninformation with the network's output. Original IBA identifies input regions\nthat have sufficient predictive information. We propose Inverse IBA to identify\nall informative regions. Thus all predictive cues for pathologies are\nhighlighted on the X-rays, a desirable property for chest X-ray diagnosis.\nMoreover, we propose Regression IBA for explaining regression models. Using\nRegression IBA we observe that a model trained on cumulative severity score\nlabels implicitly learns the severity of different X-ray regions. Finally, we\npropose Multi-layer IBA to generate higher resolution and more detailed\nattribution/saliency maps. We evaluate our methods using both human-centric\n(ground-truth-based) interpretability metrics, and human-independent feature\nimportance metrics on NIH Chest X-ray8 and BrixIA datasets. The Code is\npublicly available.\n