Investigating Image Registration Impact on Preterm Birth Classification: An Interpretable Deep Learning Approach

摘要

Deep learning algorithms have recently become the dominant trend in medical image classification. However, the decision-making rationale of convolutional neural network (CNN) classifiers can be obscure. Interpretable machine learning techniques, such as layer-wise relevance propagation (LRP), can provide a visual interpretation of these decisions. In this work, we build a 3D CNN model to classify neonatal T_2-weighted magnetic resonance (MR) scans into term or preterm. Additionally, we investigate the impact of different registration techniques applied to the image dataset on the classifier's predictions. Finally, we compute LRP 'relevance maps', which indicate each voxel's importance to the outcome of the decision. Our resulting LRP heatmaps show no visually striking differences between the different registration techniques, while also revealing anatomically plausible features for term and preterm birth.