Automatic reorientation of myocardial perfusion SPECT images

Abstract

Myocardial perfusion SPECT imaging supplies information about the blood flow in the cardiac muscle. Images must be reoriented with respect to the heart’s long axis before they can be interpreted. This step is usually performed manually, which means the process is operator-dependent and thus suffers from limited reproducibility. In this work we propose an automatic reorientation method by using a convolutional neural network in order to predict the location of the base, apex and center of the right ventricle from the acquired images. Training of the model was performed from images labeled by two different professionals. Results were compared with the inter and intra observer variabilities, verifying that the error range of the proposed method is within the same bounds as for the manual case.