HihO: accelerating artificial intelligence interpretability for medical imaging in IoT applications using hierarchical occlusion Opening the black box

摘要

In the medical imaging domain, nonlinear warping has enabled pixel-by-pixel mapping of one image dataset to a reference dataset. This co-registration of data allows for robust, pixel-wise, statistical maps to be developed in the domain, leading to new insights regarding disease mechanisms. Deep learning technologies have given way to some impressive discoveries. In some applications, deep learning algorithms have surpassed the abilities of human image readers to classify data. As long as endpoints are clearly defined, and the input data volume is large enough, deep learning networks can often converge and reach prediction, classification and segmentation with success rates as high or higher than human operators. However, machine learning and deep learning algorithms are complex. Interpretability is not always a product of the classifications performed. Visualization techniques have been developed to add a layer of interpretability. The work presented here builds on a framework for augmenting statistical findings in medical imaging workflows with machine learning results. Utilizing the framework, visualization techniques for the machine learning portion are compared in an application, and a novel, lightweight technique for machine learning visualization is proposed as a means of increasing the portability of machine learning interpretability to Internet of Things applications. The novel visualization, hierarchical occlusion, can improve time to visualization by three orders of magnitude over a traditional occlusion sensitivity algorithm.