Show abstract - PDF - Reviews

Equivariance in convolutional neural networks (CNN) has been a long-sought property, as it would ensure robustness to expected effects in the data. Convolutional filters are by nature translation-equivariant, and rotation-equivariant kernels were proposed recently. While these filters can be paired with learnable weights to form equivariant networks (E-CNN), we show here that such E-CNNs have a limited learning capacity, which makes them fragile against even slight changes in intensity distribution. This sensitivity to intensity changes presents a major challenge in medical imaging where many noise sources can randomly corrupt the data, even for consecutive scans of the same subject. Here, we propose a hybrid architecture that successively decouples intensity and spatial features: we first remove irrelevant noise in the data with a denoising CNN, and then use an E-CNN to extract robust spatial features. We demonstrate our method for motion tracking in fetal brain MRI, where it considerably outperforms standard CNNs and E-CNNs.
Hide abstract