In a development that could reshape how nuclear cardiology is practiced worldwide, researchers in Japan have shown that a type of artificial intelligence called a variational autoencoder can detect coronary artery disease in myocardial perfusion SPECT images more accurately than the standard quantitative method, and can do so without ever being told what a normal heart looks like in men versus women. The study, published in the European Journal of Nuclear Medicine and Molecular Imaging, suggests that unsupervised deep learning could one day eliminate a longstanding bottleneck in cardiac imaging: the need for painstakingly constructed, population- and sex-specific reference databases.
Myocardial perfusion SPECT, or MPS, is one of the most widely used tests in medicine for diagnosing coronary artery disease, the leading cause of death globally. The technique involves injecting a radioactive tracer that flows to the heart muscle in proportion to blood supply, then imaging the heart with a gamma camera under stress and at rest. Radiologists and automated software then compare the pattern of tracer uptake in a patient’s heart against a “normal database,” a statistical reference built from patients with a low likelihood of having the disease. When a region of the heart shows significantly less uptake than the reference, it is flagged as a perfusion defect, a hallmark of blocked coronary arteries.
The trouble with this approach is that normal perfusion patterns vary considerably between individuals. Men and women have systematically different distributions of tracer uptake, with males typically showing reduced counts in the inferior wall and females showing reduced counts in the anterior wall, largely due to differences in breast tissue attenuation and heart orientation. Adding to the complexity, normal databases must also account for ethnicity, radiopharmaceutical choice, camera technology, acquisition protocols, and reconstruction methods. A database built for Japanese patients on one gamma camera system may not translate to American patients on a cadmium-zinc telluride camera. This creates a substantial practical burden, and when an appropriate database is unavailable, quantitative analysis becomes unreliable.
The research team, led by Koichi Okuda of Hirosaki University Graduate School of Health Sciences together with colleagues at Kanazawa University and Gifu University, took a fundamentally different approach. Rather than comparing patient images to a fixed reference, they trained a variational autoencoder, or VAE, to learn the underlying distribution of normal myocardial perfusion from scratch. A VAE is a generative deep learning model consisting of three components: an encoder that compresses input images into a compact latent representation, a latent space where the data’s variability is captured as probability distributions, and a decoder that reconstructs images from points in that latent space. During training on normal images only, the VAE learns what normal hearts look like across the natural range of variation. Then, when shown a patient’s image, the decoder generates a synthetic “normal” version tailored to that patient. By comparing the patient’s original image with the generated normal image, abnormalities become visible automatically, without any explicit reference database.
To train the model, the researchers compiled 3,432 polar map images from patients without coronary artery disease, drawing on the Japanese normal databases created by the Japanese Society of Nuclear Medicine working group and additional normal datasets. Polar maps are two-dimensional “bull’s-eye” representations of the three-dimensional tracer distribution in the left ventricle, and they are the standard format for quantitative perfusion analysis. The team augmented this training data by rotating polar maps by one to five degrees, expanding the dataset to 1,771 images from females and 1,661 from males. The validation set consisted of 111 patients, of whom 43 percent had obstructive coronary artery disease, defined as at least one coronary artery with 75 percent or greater stenosis confirmed by coronary angiography.
The architecture of the VAE was carefully optimized. The researchers tested models with two to five convolutional encoder and decoder layers, each layer comprising a 2D convolution with a 4-by-4 kernel, batch normalization, and a leaky rectified linear unit activation. Hyperparameters, including the number of dense output neurons, the dropout rate, and the learning rate, were tuned using the Hyperband algorithm. The training loss combined the mean squared error of image reconstruction with Kullback-Leibler divergence, a measure of how closely the learned latent distribution matches a Gaussian prior. A three-layer model proved optimal, achieving a loss of 74.8 compared to 93.1 for the two-layer model, whose generated images were visibly degraded. Training used early stopping, halting after 100 epochs without improvement.
The performance results are striking. When the researchers measured the difference between original and VAE-generated images using six different similarity metrics, cosine similarity emerged as the best performer, achieving an area under the receiver operating characteristics curve of 0.892. The conventional approach, summed stress scores calculated with sex-segregated normal databases via QPS software, achieved an AUC of 0.814. The difference was statistically significant, with a p-value of 0.047. In practical terms, this means the VAE-based method was more accurate at distinguishing patients with obstructive coronary disease from those without, and it achieved this without requiring separate male and female reference standards.
Perhaps the most revealing findings came when the team peered inside the model’s latent space. When they mapped the training data in two dimensions, male and female data points clustered separately, with centroids separated by a stable distance across model architectures, even though some overlap existed. Decoding individual points across the latent space revealed that the model had spontaneously learned sex-specific perfusion patterns: one corner of the latent space generated uniformly distributed perfusion characteristic of normal males, while another generated the reduced inferior uptake typical of normal females. The VAE had disentangled the sources of normal variation without any explicit instruction, effectively building its own internal normal databases for each sex.
The clinical implications are considerable. First, the approach could reduce the enormous logistical effort of creating and validating normal databases for every combination of population, camera, protocol, and tracer. A single VAE trained on data from a given imaging setting could internally adapt to individual characteristics, including sex, rather than forcing a choice between male and female databases. Second, the generative nature of the model makes its decisions interpretable. Clinicians can see the synthetic normal image the model generated for a given patient, along with a subtracted image highlighting the suspicious region. In one example from the study, the model generated a synthetic normal perfusion pattern from a patient with anterior ischemia, and the subtracted image clearly delineated the ischemic territory in the apex to anterior wall, corresponding to the left anterior descending artery. This kind of visual explanation is critical for building trust in AI tools.
The authors caution that several questions remain before clinical deployment. The study used polar map images rather than raw SPECT slices, so some spatial information from the myocardial walls is lost in the projection; a model trained directly on three-dimensional SPECT data might perform differently. All images were acquired on conventional Anger gamma cameras with 360-degree acquisition and filtered back-projection reconstruction, so the approach’s validity with modern cadmium-zinc telluride cameras, 180-degree protocols, or iterative reconstruction remains to be tested. The applicability to iodine-123 tracers, which have different photon energies and scatter properties, is also unexplored. Furthermore, the sample size for training was not formally optimized, and the clinical validity of the synthetic normal images warrants further study.
Even with these caveats, the study represents a notable shift in how machine learning is applied to nuclear cardiology. Previous deep learning applications in MPS have largely focused on classification or prediction, training networks to output diagnostic labels directly. The novelty here lies in image generation: the network creates a personalized normal reference, and diagnosis follows from a straightforward comparison. This unsupervised paradigm sidesteps the need for disease-labeled training data, which is often scarce and expensive to obtain, and offers a path toward AI tools that generalize more gracefully across populations and imaging systems. For a field that has spent decades refining normal databases one variable at a time, the idea that a single generative model can learn those variations on its own is a compelling glimpse of where quantitative cardiac imaging is headed.
Subject of Research: Detection of coronary artery disease in myocardial perfusion SPECT images using a convolutional variational autoencoder trained on normal perfusion polar maps
Subject of Research: Medicine
Article Title: Variational autoencoders can detect coronary artery disease in myocardial perfusion SPECT images
Article References: Okuda, K., Nakajima, K., Hara, T., Yoneyama, H., & Kinuya, S. (2026). Variational autoencoders can detect coronary artery disease in myocardial perfusion SPECT images. European Journal of Nuclear Medicine and Molecular Imaging. https://doi.org/10.1007/s00259-026-08161-1
Image Credits: AI Generated
DOI: 10.1007/s00259-026-08161-1
Keywords: Variational autoencoder, Myocardial perfusion, SPECT, Coronary artery disease, Anomaly detection, Deep learning, Polar maps, Cosine similarity, Latent space, Normal database, Sex-specific perfusion, Nuclear cardiology
Cite Scienmag News
APA MLA Chicago
Ophelia Keating. (September 4, 2026). Variational autoencoders detect coronary artery disease in SPECT images. Scienmag. https://scienmag.com/variational-autoencoders-detect-coronary-artery-disease-in-spect-images/
Ophelia Keating. “Variational autoencoders detect coronary artery disease in SPECT images.” Scienmag, 4 September 2026, https://scienmag.com/variational-autoencoders-detect-coronary-artery-disease-in-spect-images/. Accessed 4 September 2026.
Ophelia Keating. “Variational autoencoders detect coronary artery disease in SPECT images.” Scienmag. September 4, 2026. https://scienmag.com/variational-autoencoders-detect-coronary-artery-disease-in-spect-images/
Copy citation Download RIS
Tags: advancements in nuclear medicine diagnosticsAI-based analysis of SPECT imagesAI-based diagnosis of cardiac conditionsAI-driven improvements in myocardial perfusion testingautomated analysis of SPECT imagesautomated diagnosis of coronary artery diseasechallenges in traditional cardiac imaging methodscoronary artery disease detectiondeep learning models for cardiac image interpretationdeep learning models for coronary artery diseaseinnovative AI methods in nuclear medicinemyocardial perfusion SPECT imagingnuclear cardiology advancementsovercoming limitations of traditional cardiac imaging databasesradioactive tracer imaging for heart healthreduction of population-specific reference requirementssex-specific cardiac imaging referencessex-specific reference databases in medical imagingunsupervised deep learning in cardiac imagingunsupervised deep learning in cardiologyvariational autoencoders in medical diagnosisvariational autoencoders in nuclear cardiology
