SCMR 22nd Annual Scientific Sessions
Cardiac T1-mapping is increasingly used for advanced myocardial tissue characterisation, and machine learning promises to automate cumbersome post-processing for translation into real-world practice. Deep Learning (DL) algorithms are highly suitable for automation of medical image analysis, but its unaccountable ‘black box’ behaviour remains a chief concern for clinical applications. We present a DL method with visualisation techniques, to shed light into the machine perception process and demonstrate its application in automated detection of motion artefacts in cardiac T1-mapping, which is a critical quality assurance step for clinically-robust T1 determinations.
Approximate 5000 ShMOLLI T1-maps  from the UK Biobank were scored for motion by an observer with over 10 years’ experience in CMR image processing. The dataset was randomly split into 80% training data and 20% validation data. A DL architecture utilising residual neural network (ResNet)  blocks was developed, consisting of two ResNet streams extracting features from the seven raw T1-weighted images and R2 quality control map, respectively. Features from the streams were concatenated for the final prediction of motion artefacts. A visualisation module was developed to offer human observers additional insight into the machine perception process of individual input images.
This DL method demonstrated 88% agreement with an experienced human observer in motion detection on the validation dataset (85% specificity, 91% sensitivity). With the R2 map stream added to the neural network, agreement, specificity, and sensitivity improved to 91%, 88%, and 92% respectively, indicating the effectiveness of the multi-stream network and incremental usefulness of the composite information contained in R2 map for artefact detection. Saliency visualisation revealed that the decision-making process of the network focuses on myocardial regions in the raw images, and the dark regions in the R2 map for motion detection (Figure 1). The generated heat maps (Figure1, e-h) demonstrated a good correlation with the regions affected by motion (red arrows). Attention maps in Figure 2 show that machine focused on gastrointestinal rather than cardiac motion artefacts, which helped to identify the source of disagreement with experts.
Deep learning saliency visualisation allows human operators additional insight into the deep learning ‘black box’ in diagnostic medical imaging. Attention heat maps provide a traceable record of the perception process of the machine, offering additional control measures to ensure accountability required for clinical applications. Combining attention maps with accurate myocardial segmentation may help automatic rejection of artefacts that carry no clinical relevance.