Premium
Deep learning models for generating diagnostic explanations
Author(s) -
Dyrba Martin,
Pallath Arjun H.,
Marzban Eman N.,
Teipel Stefan J.
Publication year - 2020
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.037353
Subject(s) - relevance (law) , interpretability , computer science , convolutional neural network , artificial intelligence , visualization , deep learning , machine learning , neuroimaging , pattern recognition (psychology) , neuroscience , psychology , law , political science
Background Although machine learning approaches achieve high diagnostic accuracy when detecting Alzheimer’s disease (AD) based on MRI scans, they are rarely applied in clinical studies due to a lack of suitable methods for model comprehensibility and interpretability. Recent advances in convolutional neural networks (CNN) visualization algorithms and methods for generating textual explanations may fill this gap. Here, we will introduce these novel approaches, present first prototypical software using these techniques, and outline the potential of such tools for clinical studies. Method We implemented a CNN model trained on 662 T1‐weighted MRI scans obtained from ADNI. Various CNN visualization algorithms were compared, which generate relevance maps indicating the contribution of individual image areas for detecting AD. Further, we created an interactive web application to display them and interactively change visualization parameters. For evaluation, we compared relevance scores in the hippocampus region with hippocampal volume. Result The CNN model reached a cross‐validated area under the curve (AUC) of 0.93±0.06 for AD vs. controls and 0.74±0.09. The layer‐wise relevance propagation (LRP) algorithm was found to provide most informative and robust results (Fig. 1). The web app is shown in Fig. 2. Hippocampal volume and relevance scores of the hippocampus region‐of‐interest correlated with r= ‐0.79 (Fig. 3). Additionally, relevance maps highlighted various other image regions with local cortical and subcortical atrophy. Conclusion CNN relevance maps may support clinicians in their MRI examinations. CNN models for generating textual explanations may offer additional information, but are not yet available due to the vast amount of labeled training samples required for training.