This technology offers an automated system and method to preprocess M
PI and use DL models to predict significant coronary artery disease (CAD).
The preprocessing includes automated segmentation of the myocardial re
gion, automatic registration to a template, activity normalization, and sph
erical coordinate transformation. These techniques enable three-dimensio
nal (3D) image analysis and prediction by the deep learning (DL) model, w
hich overcomes the orientation problem for 3D myocardial structures with
convolutional neural networks. Validated through cross-validation with ne
arly a thousand clinical scans and external validation with nearly a thousa
nd additional cases, the technology surpasses traditional total perfusion d
eficit (TPD) prediction, achieving an area under the receiver operating cha
racteristic curve of 0.844 compared to TPD's 0.759, while maintaining th
e same specificity and increasing sensitivity from 76.6% to 83.8%. Importa
ntly, this technology does not relay on a normal database, avoiding the ti
me-consuming step of collecting data from healthy subjects.