Incorporation of prior knowledge about organ shape and location is key toimprove performance of image analysis approaches. In particular, priors can beuseful in cases where images are corrupted and contain artefacts due tolimitations in image acquisition. The highly constrained nature of anatomicalobjects can be well captured with learning based techniques. However, in mostrecent and promising techniques such as CNN based segmentation it is notobvious how to incorporate such prior knowledge. State-of-the-art methodsoperate as pixel-wise classifiers where the training objectives do notincorporate the structure and inter-dependencies of the output. To overcomethis limitation, we propose a generic training strategy that incorporatesanatomical prior knowledge into CNNs through a new regularisation model, whichis trained end-to-end. The new framework encourages models to follow the globalanatomical properties of the underlying anatomy (e.g. shape, label structure)via learned non-linear representations of the shape. We show that the proposedapproach can be easily adapted to different analysis tasks (e.g. imageenhancement, segmentation) and improve the prediction accuracy of thestate-of-the-art models. The applicability of our approach is shown onmulti-modal cardiac datasets and public benchmarks. Additionally, wedemonstrate how the learned deep models of 3D shapes can be interpreted andused as biomarkers for classification of cardiac pathologies.
展开▼
