Afterimages: integration of diagnostic information through Bayesian enhancement of scintigraphic images

Although the diagnostic accuracy of myocardial perfusion scintigraphy can be improved by additional consideration of clinical and exercise information, multivariate prediction models are infrequently used for this purpose in the clinical setting. We therefore developed a Bayesian algorithm that instead transforms the scintigraphic image itself, modifying defect contrast as a function of the pretest likelihood of coronary artery disease. The algorithm was tested in computer simulations of myocardial perfusion scintigraphy with data from 378 patients (166 from California and 212 from West Virginia) who underwent planar exercise thallium-201 scintigraphy and coronary angiography. Images were interpreted before and after enhancement by eight readers (four at each medical center) with different training orientations (internist, radiologist, cardiologist, nuclear cardiologist, and nuclear medicine technologist) who used a four-point score (from 0, normal to 3, severe defect). Accuracy was quantified as area under a receiver-operating characteristic (ROC) curve. Improvements in accuracy obtained by the algorithm were compared to those provided by multiple logistic regression. Overall, Bayesian enhancement increased ROC area from 0.63 +/- 0.04 to 0.71 +/- 0.04 (p < 0.01). The improvement was consistent for all 16 reading sets (eight readers multiplied by two patient populations; p < 0.05). In comparison, multiple logistic regression increased ROC area from 0.63 +/- 0.04 to 0.79 +/- 0.03 (p < 0.01), outperforming interpretation of the enhanced images in 13 of the 16 reading sets. Bayesian enhancement improves diagnostic accuracy of conventional scintigraphic image interpretation. The improvement is stable across individuals, training orientations, and patient populations. Although this approach is not as accurate as multiple logistic regression, it may be more practical for widespread clinical application.