Variational autoencoders can detect coronary artery disease in myocardial perfusion SPECT images.

Abstract

Variational autoencoders (VAEs) can detect unsupervised anomalies in medical images. We aimed to determine whether a VAE could automatically detect anomalies in myocardial perfusion SPECT (MPS) images. We then interpreted the training and generative processes of the VAE. We trained a VAE on polar maps to learn the variability of normal myocardial perfusion profiles in latent space. The training dataset comprised 3,432 polar map images (49% male) without coronary artery disease (CAD) and the validation set contained 111 polar map images (59% male), among whom, (43% had obstructive CAD). This latent representation allowed the VAE to reconstruct synthetic normal images that served as references to detect anomalies. Abnormal myocardial perfusion was automatically identified by measuring differences between generated and original images using cosine similarity. We also summed stress scores (SSS) using sex-segregated normal databases. The trained VAE generated reference normal polar maps, regardless of normal and defect profiles and sex differences. The areas under the receiver-operating characteristics curve to detect CAD were 0.89 ± 0.041 for cosine similarity and 0.81 ± 0.030 for SSS (p = 0.047). The VAE learned the characteristics of normal male and female myocardial perfusion distributions in the latent space. This VAE-based approach improved CAD detection in MPS images and eliminated the need to determine sex-specific myocardial perfusion characteristics.