A Localisation Study of Deep Learning Models for Chest X-ray Image Classification

Deep learning models have demonstrated superhuman performance in a multitude of image classification tasks, including the classification of chest X-ray images. Despite this, medical professionals are reluctant to embrace these models in clinical settings due to a lack of interpretability, citing being able to visualise the image areas contributing most to a model's predictions as one of the best ways to establish trust. To aid the discussion of their suitability for real-world use, in this work, we attempt to address this issue by conducting a localisation study of two state-of-the-art deep learning models for chest X-ray image classification, ResNet-38-large-meta and CheXNet, on a set of 984 radiologist annotated X-ray images from the publicly available ChestX-ray14 dataset. We do this by applying and comparing several state-of-the-art visualisation methods, combined with a novel dynamic thresholding approach for generating bounding boxes, which we show to outperform the static thresholding method used by similar localisation studies in the literature. Results also seem to indicate that localisation quality is more sensitive to the choice of thresholding scheme than the visualisation method used, and that a high discriminative ability as measured by classification performance is not necessarily sufficient for models to produce useful and accurate localisations.