Understanding the role of saliency maps for biomarker research in 3D medical imaging classification

Convolutional Neural Networks (CNNs) have achieved promising success on 3D medical imaging analysis, for instance classification of glioma tumours on Magnetic Resonance (MR) images. However, CNN are considered to be 'black boxes' due to their nontransparent characteristics in the learning process. To exposing the intrinsic actuating patterns of CNN, researchers have proposed a serial of explanation methods for translating CNN's decision mechanism into visualised anatomical representation. In the 3D medical imaging-based field, the interpretation is obtained via saliency maps, which present the contribution of input voxels associated with the network outputs. Therefore, it is significant to understand if saliency maps can be considered as potential biomarkers by providing reliable anatomical information. In the paper, we conducted a validation analysis to measure the robustness of saliency maps with respect to various properties in 3D medical imaging classification tasks. Furthermore, we proposed a novel method to generate a synthetic dataset by mimicking the appearance and structure of 3D medical imaging with ground truth information for easily estimating the accuracy of saliency maps under the limitation of annotations provided in the classification tasks. Our experiment results demonstrate all selected explanation methods fail in at least one measurement. Researchers should be critically careful to take advantage of saliency maps as biomarkers in 3D medical imaging classification tasks.