Comparing radiologists' gaze and saliency maps generated by interpretability methods for chest x-rays
This work addresses the interpretability of medical image analysis models for clinicians, but it is incremental as it applies existing methods to a new dataset with comparative evaluation.
The study compared saliency maps from interpretability methods (Grad-CAM and attention-gated models) with radiologists' gaze data on chest x-rays using shuffled metrics to avoid biases, finding that Grad-CAM achieved scores comparable to an interobserver baseline in one metric, indicating potential to mimic radiologist attention.
The interpretability of medical image analysis models is considered a key research field. We use a dataset of eye-tracking data from five radiologists to compare the outputs of interpretability methods and the heatmaps representing where radiologists looked. We conduct a class-independent analysis of the saliency maps generated by two methods selected from the literature: Grad-CAM and attention maps from an attention-gated model. For the comparison, we use shuffled metrics, which avoid biases from fixation locations. We achieve scores comparable to an interobserver baseline in one shuffled metric, highlighting the potential of saliency maps from Grad-CAM to mimic a radiologist's attention over an image. We also divide the dataset into subsets to evaluate in which cases similarities are higher.