Interpretable deep generative model in myocardial perfusion SPECT images based on the creation of patient-matched normal maps

Background Although conventional quantitative method have used statistical normal databases for myocardial perfusion imaging, the potential role of generative deep learning has not been investigated. Purpose We sought to interpret training and generative processes of myocardial perfusion SPECT images in a variational autoencoder (VAE) (1) by comparing conventional quantitative analysis using normal maps. Methods A total of 3,432 polar map images without known coronary artery disease (CAD) were used as a training dataset, and 111 polar map images (43% obstructive CAD) were also used as a validation dataset. The VAE was trained to detect perfusion abnormalities using normal images, and consisted of an encoder, latent space, and a decoder (Fig.1). During the training process using the training dataset, the features of input normal images were characterized by the encoder in the latent space, where the probability of data points was represented as the Gaussian distribution. In the generative process, whatever perfusion distribution a patient had, the decoder sampled the patient’s matched perfusion distribution from the latent space, and generated a "normal" polar map image fitted for individual patients. Perfusion abnormality was detected by computing a cosine similarity (2) between the input and generated images. Results Distributions of normal myocardial perfusion for males and females were represented in the latent space (Fig.2a). Although male and female data points were overlapped, the distance between centroids of the points was 0.57. When normal polar maps were generated from the latent space, uniform myocardial perfusion was exhibited in the upper left corner, and decreased myocardial perfusion at the inferior was exhibited in the lower right corner (Fig. 2b). Uniform and decreased myocardial perfusion agreed with the characteristics of male and female myocardial perfusion, respectively. Trained VAE model was able to generate a normal polar map image from abnormal perfusion in the patient with CAD (Fig.2c). The areas under the receiver-operating characteristic curve for CAD detection were 0.89 ± 0.041 for cosine similarity and 0.81 ± 0.030 for summed stress score (p = 0.047). Conclusions The VAE automatically learned the characteristics of normal male and female myocardial perfusion distributions in the latent space, and estimated sex-specific normal perfusion distributions in generative polar map images. The VAE model has the potential to improve the detection of CAD in myocardial perfusion SPECT images.Fig 1.Structure of VAE model Fig 2.