Manifold learning for patient position detection in MRI

Magnetic resonance imaging is performed without ionizing radiation, however, the applied radio frequency power leads to heating, which is dependent on the body part being imaged. Determining the patient position in the scanner allows to better monitor the absorbed power and therefore optimize the image acquisition. Low-resolution images, acquired during the initial placement of the patient in the scanner, are exploited for detecting the patient position. We use Laplacian eigenmaps, a manifold learning technique, to learn the low-dimensional manifold embedded in the high-dimensional image space. Our experiments clearly show that the presumption of the slices lying on a low dimensional manifold is justified and that the proposed integration of neighborhood slices and image normalization improves the method. We obtain very good classification results with a nearest neighbor classifier operating on the low-dimensional embedding.