Development of artificial intelligence algorithms for optimizing patient positioning for cardiac SPECT in nuclear medicine

Cardiac SPECT is a non-invasive tomographic nuclear medicine technique which allows to perform a myocardial perfusion imaging (MPI). During this procedure, a specifically designed radiotracer is administered to a patient and it concentrates in the heart tissues. The 3D distribution of the radiopharmaceutical activity inside the heart is reconstructed by collecting enough 2D projection images around the patient. This technique is a powerful tool for coronary artery disease (CAD) diagnosis. Comparing two sets of images, one acquired after a stress test and another in a rest state, a SPECT MPI study highlights possible perfusion defects, which may result in a myocardial ischemia or a myocardial infarction. D-SPECT is a cardiac-dedicated system which fits the left side of the patient’s chest with its curved geometry. This camera shows a much greater sensitivity, a higher reconstructed resolution and an improved energy resolution, compared to the conventional systems. The D-SPECT image quality may be degraded by several factors. Detector and patient chair positions are not defined in advance, so that the patient-detector configuration depends on the operator and on the patient characteristics. In addition, the cardiac region should be identified on the pre-scan in order to focus the final acquisition on the heart. This procedure is strongly operator-dependent. This work is aimed at developing artificial intelligence algorithms which facilitate these procedures, providing the operator with indications on the patient positioning and the heart location in the preliminary scan. 711 scans related to 158 different patients were collected, using a D-SPECT system at the Nuclear Medicine Department of the San Giovanni Battista University Hospital of Turin. A graphical user interface (GUI) was developed to allow experts to select the heart position in the pre-scan and to indicate possible corrections to the patient-detector configuration. Specifically, the technicians denoted the pre-scans for which the head detector or the patient chair are not in the right position. They also labelled the acquisitions which presented a great extra-cardiac activity, suggesting the patient to drink cold water. Support vector machine (SVM) models were trained by using the collected information. A classifier for detecting the heart position was developed, showing a sensitivity up to 96%. In order to assist the operator during the diagnostic test, two classifiers were developed. One has the aim to distinguish among three different pre-scan categories: raise the chair, lower the chair or leave the chair as it is for the semi-upright position; move the detector head towards the patient's feet, towards the patient’s head or leave the detector head as it is. The second classifier is useful to discern between two other classes: make the patient drink or not drink cold water, depending on the external activity. On average, these algorithms achieved an accuracy of 75%. Even if these algorithms are not completely developed, they are surely a good starting point to make the diagnostic work easier for the operators.