What Does Deep Learning See? Insights From a Classifier Trained to Predict Contrast Enhancement Phase From CT Images.
AJR Am J Roentgenol. 2018 Dec;211(6):1184-1193. doi: 10.2214/AJR.18.20331. Epub 2018 Nov 7.
Philbrick KA1, Yoshida K1, Inoue D1, Akkus Z1, Kline TL1, Weston AD1, Korfiatis P1, Takahashi N1, Erickson BJ1.
OBJECTIVE: Deep learning has shown great promise for improving medical image classification tasks. However, knowing what aspects of an image the deep learning system uses or, in a manner of speaking, sees to make its prediction is difficult.
MATERIALS AND METHODS: Within a radiologic imaging context, we investigated the utility of methods designed to identify features within images on which deep learning activates. In this study, we developed a classifier to identify contrast enhancement phase from whole-slice CT data. We then used this classifier as an easily interpretable system to explore the utility of class activation map (CAMs), gradient-weighted class activation maps (Grad-CAMs), saliency maps, guided backpropagation maps, and the saliency activation map, a novel map reported here, to identify image features the model used when performing prediction.
RESULTS: All techniques identified voxels within imaging that the classifier used. SAMs had greater specificity than did guided backpropagation maps, CAMs, and Grad-CAMs at identifying voxels within imaging that the model used to perform prediction. At shallow network layers, SAMs had greater specificity than Grad-CAMs at identifying input voxels that the layers within the model used to perform prediction.
CONCLUSION: As a whole, voxel-level visualizations and visualizations of the imaging features that activate shallow network layers are powerful techniques to identify features that deep learning models use when performing prediction.