Automated interpretation of gated-SPECT. A new approach to integrate the analysis of left ventricular perfusion and function

Abstract

The general aim of the thesis was to develop and validate an automated decision support system for the interpretation of gated-SPECT myocardial perfusion images. Myocardial perfusion scintigraphy is a widely used method for the diagnosis and evaluation of patients with suspected or known coronary artery disease. The interpretation of these images can be difficult, and errors in interpretation may lead to serious mistakes in the treatment of patients. Thus, tools have been developed to assist physicians with the aim of improving accuracy and reducing variability between readers.The cornerstones of our decision support system are image processing techniques, artificial neural networks and databases of classified diagnostic images. We developed a new image processing method for quantification of cardiac function, denoted CAFU, which is based on the active shape algorithm and a heart-shaped model. Artificial neural networks that were inspired by the structure and function of biological neural networks such as the human brain were used to interpret the images based on CAFU measurements of left ventricular perfusion and function. The image processing and artificial neural network techniques are data-driven, i.e. they rely on a large number of relevant examples. Thus, large databases of myocardial perfusion images were developed.The decision support system was evaluated in a hospital separate from the location at which it was trained, and compared with an automated quantification software package. The system showed a performance regarding detection of myocardial infarction, measured as areas under receiver operating characteristics curves, of between 91% and 99% in the test material. The system also showed significantly higher specificities (95%) than the quantification software package (50% and 74%, respectively), compared at a sensitivity of 91%.In conclusion, our decision support system presents interpretations more similar to those of experienced clinicians than to those of a conventional automated quantification software package. This study shows the feasibility of disseminating the expertise of experienced clinicians to less experienced physicians by the use of artificial neural networks.