Assessing the (Un)Trustworthiness of Saliency Maps for Localizing\n Abnormalities in Medical Imaging

Abstract

Saliency maps have become a widely used method to make deep learning models\nmore interpretable by providing post-hoc explanations of classifiers through\nidentification of the most pertinent areas of the input medical image. They are\nincreasingly being used in medical imaging to provide clinically plausible\nexplanations for the decisions the neural network makes. However, the utility\nand robustness of these visualization maps has not yet been rigorously examined\nin the context of medical imaging. We posit that trustworthiness in this\ncontext requires 1) localization utility, 2) sensitivity to model weight\nrandomization, 3) repeatability, and 4) reproducibility. Using the localization\ninformation available in two large public radiology datasets, we quantify the\nperformance of eight commonly used saliency map approaches for the above\ncriteria using area under the precision-recall curves (AUPRC) and structural\nsimilarity index (SSIM), comparing their performance to various baseline\nmeasures. Using our framework to quantify the trustworthiness of saliency maps,\nwe show that all eight saliency map techniques fail at least one of the\ncriteria and are, in most cases, less trustworthy when compared to the\nbaselines. We suggest that their usage in the high-risk domain of medical\nimaging warrants additional scrutiny and recommend that detection or\nsegmentation models be used if localization is the desired output of the\nnetwork. Additionally, to promote reproducibility of our findings, we provide\nthe code we used for all tests performed in this work at this link:\nhttps://github.com/QTIM-Lab/Assessing-Saliency-Maps.\n

Ähnliche Arbeiten