Interpretation of SPECT wall motion with deep learning

Abstract

ObjectivesWe sought to develop a novel deep learning (DL) workflow to interpret single-photon emission computed tomography (SPECT) wall motion. BackgroundWall motion assessment with SPECT is limited by image temporal and spatial resolution. Visual interpretation of wall motion can be subjective and prone to error. Artificial intelligence (AI) may improve accuracy of wall motion assessment. MethodsA total of 1038 patients undergoing rest electrocardiogram (ECG)-gated SPECT and echocardiography were included. Using echocardiography as truth, a DL-model (DL-model 1) was trained to predict the probability of abnormal wall motion. Of the 1038 patients, 317 were used to train a DL-model (DL-model 2) to assess regional wall motion. A 10-fold cross-validation was adopted. Diagnostic performance of DL was compared with human readers and quantitative parameters. ResultsThe area under the receiver operating characteristic curve (AUC) and accuracy (ACC) of DL model (AUC: .82 [95% CI: .79-.85]; ACC: .88) were higher than human (AUC: .77 [95% CI: .73-.81]; ACC: .82; P < .001) and quantitative parameter (AUC: .74 [95% CI: .66-.81]; ACC: .78; P < .05). The net reclassification index (NRI) was 7.7%. The AUC and accuracy of DL model for per-segment and per-vessel territory diagnosis were also higher than human reader. The DL model generated results within 30 seconds with operable guided user interface (GUI) and therefore could provide preliminary interpretation. ConclusionsDL can be used to improve interpretation of rest SPECT wall motion as compared with current human readers and quantitative parameter diagnosis.