Automatic view classification for cardiac MRI

View classification for cardiac MR images is a new topic in medical image analysis, and can support efficient content-based filtering, browsing, and retrieval. The major difficulty lies in large variability in image appearance caused by various acquisition protocols, heart phases and disease conditions. We propose a collaborative learning approach that exploits statistical dependencies at three levels: the local image patch level, the parts-whole level, and the anatomical level. Specifically, at the local image patch level, we make redundant use of the training images in a spatial manner; At the parts-whole level, we model the relationship among the detected landmarks by a sparse configuration method to remove erroneously detected landmarks; At the anatomical level, we train a view classifier based on the DICOM header as another information source, which is then optimally incorporated in a Bayesian way as a prior. We compare the approach with various integrations of state-of-art methods. Large-scale experiments on real-world clinical datasets - over ten thousand unseen images, many with severe diseases - show that the approach is highly robust, outperforming other integrated approaches, and can achieve a 97.6% classification accuracy.