Myocardial perfusion scintigraphy image database

Abstract

This database provides a collection of myocardial perfusion scintigraphy images in DICOM format with all metadata and segmentations (masks) in NIfTI format. The images were obtained from patients undergoing scintigraphy examinations to investigate cardiac conditions such as ischemia and myocardial infarction. The dataset encompasses a diversity of clinical cases, including various perfusion patterns and underlying cardiac conditions. All images have been properly anonymized, and the age range of the patients is from 20 to 90 years. This database represents a valuable source of information for researchers and healthcare professionals interested in the analysis and diagnosis of cardiac diseases. Moreover, it serves as a foundation for the development and validation of image processing algorithms and artificial intelligence techniques applied to cardiovascular medicine. Available for free on the PhysioNet platform, its aim is to promote collaboration and advance research in nuclear cardiology and cardiovascular medicine, while ensuring the replicability of studies.

Background

Myocardial perfusion scintigraphy (MPS), also known as myocardial perfusion imaging (MPI), is a noninvasive nuclear imaging modality used to assess myocardial blood flow under rest and stress conditions. It plays a central role in the diagnosis, risk stratification, and follow-up of patients with suspected or confirmed coronary artery disease (CAD). By visualizing perfusion patterns in the left ventricular myocardium, MPS allows for the detection of ischemia, infarction, and viability, contributing to the personalization of therapeutic decisions.

SPECT-based MPS with technetium-99m–labeled radiotracers remains the most widely used technique in nuclear cardiology due to its accessibility, cost-effectiveness, and extensive validation in clinical trials. According to guidelines from the American Society of Nuclear Cardiology (ASNC), MPS remains a first-line imaging option in several clinical scenarios, particularly in patients with intermediate pretest likelihood of CAD or those unable to perform stress echocardiography [1].

However, the interpretation of SPECT images is subject to several limitations. Attenuation artifacts, motion-related distortions, and variability in segmentation protocols may reduce diagnostic confidence and consistency. Manual delineation of the left ventricular wall, when required, is time-consuming and prone to inter-observer differences. These limitations underscore the need for computational tools that enhance objectivity, reproducibility, and efficiency in image post-processing.

Artificial intelligence (AI) methods—particularly convolutional neural networks—have emerged as promising solutions for addressing these challenges. Recent research using this dataset has demonstrated that nnU-Net, a self-configuring deep learning framework, can perform accurate segmentation of the left ventricle in SPECT-MPI images, even in the presence of artifacts. Moreover, the dataset has also supported investigations into the automated identification of diagnostic artifacts, reinforcing the potential of AI to support quality assurance and decision-making in nuclear medicine workflows [2, 3].

The primary motivation for sharing this dataset is to provide the research community with access to real-world clinical images that can be used to develop, train, and validate algorithms for segmentation, artifact detection, and diagnostic automation in cardiac imaging. The dataset includes anonymized SPECT images in both DICOM (raw) and NIfTI (processed) formats, along with ground-truth segmentation masks for selected cases. Its structure and metadata are compatible with medical image processing pipelines, enabling reproducible experimentation and model benchmarking in AI-based nuclear cardiology.

Methods

The procedure commences with the myocardial examination, wherein images are acquired using the Discovery NM 530c system from GE Healthcare Ltd., a multi-pinhole gamma camera equipped with nineteen Cadmium Zinc Telluride (CZT) semiconductor detectors arranged in a 32 × 32 px matrix arc configuration, ensuring all pinholes are focused on the heart. The images were acquired for a prospective study and include all patients referred for myocardial perfusion scintigraphy (MPS), regardless of coronary artery disease history. All participants voluntarily signed an Informed Consent Form (ICF), authorizing the use of their images after examination and report completion.

All images refer to rest myocardial perfusion SPECT scans. The examinations were conducted following a standard one-day protocol, with patients in a supine position. A weight-adjusted dose of ^99mTc-MIBI was administered, and rest-phase images were acquired one hour post-injection using the Discovery NM 530c system. Acquisition parameters included a 6-minute scan duration, a 20% energy window centered at 140 keV, a pixel size of 2.46 × 2.46 mm, and a zoom factor of 1. The images were reconstructed using the Myovation for Alcyone software (GE Healthcare Ltd.) with the Maximum Likelihood-Expectation Maximization (MLEM) algorithm.

The dataset was collected at Hospital Israelita Albert Einstein, São Paulo, Brazil, between February and March 2023. The sole exclusion criterion was non-consent; thus, rest-phase images from all consenting patients were included. Although diagnostic reports were available, only DICOM image copies were used for this study, with originals stored in the institutional PACS. Demographic data were recorded in REDCap, and no sensitive clinical data were accessed by the research team. The dataset comprises 83 patients aged between 20 and 90 years, with 88% male participants. The dataset comprises only resting myocardial perfusion SPECT scans, acquired under a standardized one-day protocol. No stress-induced acquisitions were performed. All studies followed consistent parameters using a dedicated gamma camera system, and were anonymized in compliance with the DICOM PS3.15 Annex E (Basic Application Level Confidentiality Profile).

Data Description

The dataset contains myocardial perfusion scintigraphy (MPS) images from 83 unique patients, acquired using a CZT-based gamma camera (Discovery NM 530c, GE Healthcare). All images were obtained under rest-only protocols using a standardized one-day acquisition procedure with 99mTc-MIBI. The dataset is anonymized and provided in two formats: DICOM (original input) and NIfTI (preprocessed volumes and segmentation masks, when available).

File Organization

Files are named using anonymized identification codes. Each code starts with the prefix 1.2.840.4267.32, followed by a randomly generated numeric suffix of 35–40 digits. For each patient, the DICOM and NIfTI files share the same base identifier. Segmentation masks, when available, are distinguished by the _mask.nii.gz suffix (e.g., 1.2.840.4267.32.<random-suffix>.nii.gz and 1.2.840.4267.32.<random-suffix>_mask.nii.gz).

The dataset is organized into three main directories:

• DICOM/ — raw SPECT images in DICOM format (*.dcm).
• NIfTI/ — converted images in NIfTI format (*.nii.gz).
• segmentation_masks/ — binary masks manually generated by clinical experts, where label 1 denotes the myocardial wall and 0 denotes the background.

Demographic and Metadata

Demographic information is available in the file demographics.csv, which includes patient_id, sex (M/F), age (years), and BMI (when available). Each row corresponds to a single study and links directly to the imaging data.

Technical metadata, extracted from DICOM headers, are consistent across all studies and include:

• Modality: NM (Nuclear Medicine)
• Manufacturer: GE Medical Systems
• Scanner model: Discovery NM 530c
• Collimator type: Pinhole (MPH_27_45)
• Radiopharmaceutical: 99mTc-MIBI
• Energy window: 126.45–154.55 keV
• Patient position: Supine, feet-first
• Image resolution: 70×70 pixels, with 4.0×4.0 mm² pixel spacing
• Slice thickness: 4.0 mm, 50 slices per volume
• Bit depth: 16-bit
• Window center/width: 16383/32767
• Deidentification protocol: Basic Application Confidentiality Profile (DICOM codes 113100, 113101, 113107)

Across the dataset, there are 101 studies and 103 DICOM images, as two studies were not segmented. This results in 101 NIfTI volumes (with masks) and 103 original DICOM files. All NIfTI files were generated directly from the DICOM images using a standardized preprocessing process.

Descriptive Statistics

• Total studies: 83
• Sex distribution: 88% male, 12% female
• Age range: 20–90 years
• Diagnosis labels: Not provided
• Confirmed ischemia/necrosis: Not labeled in this release

These descriptive statistics are included here for immediate reference.

Notes on Reuse

This dataset is intended to support research in segmentation and preprocessing methods in nuclear cardiology. Future releases may incorporate diagnostic labels or artifact annotations, pending ethical approval and enhanced anonymization.

Further details, including the complete description of the dataset structure and metadata, are available in the file data_description.txt included in the project repository.

Usage Notes

This dataset is intended to support a wide range of research activities in nuclear cardiology and medical image computing. Potential applications include:

• Left ventricular wall segmentation using fully supervised learning models such as nnU-Net.
• Automated detection of ischemia through radiomics or deep learning algorithms applied to perfusion patterns.
• Artifact detection and quality control, including motion, attenuation, and reconstruction artifacts, via image analysis or neural networks.
• Data harmonization and preprocessing workflows, such as image normalization, reorientation, and intensity standardization for SPECT data.
• Educational use in teaching medical image processing, segmentation, and quality assurance in nuclear medicine.

All images are anonymized and provided in both raw DICOM and processed NIfTI formats, facilitating their integration into established toolchains. The NIfTI volumes and segmentation masks (when available) are compatible with frameworks such as nnU-Net, 3D Slicer, ITK-SNAP, and Python libraries including nibabel, SimpleITK, and pydicom.

To support reproducibility and accelerate adoption, we provide a companion software capsule with ready-to-run nnU-Net scripts and configuration files. The automatic segmentation process was conducted using nnU-Net 2D (fold 0), a widely adopted self-configuring deep learning framework for biomedical image segmentation. The training and inference were executed within a Docker environment for ease of deployment.

Limitations

The dataset includes only rest-phase acquisitions (no stress-phase images). All data were collected from a single institution using a specific CZT-based camera (GE Discovery NM 530c), which may limit generalizability. There are no diagnostic outcome labels (e.g., ischemia, infarction) or clinical reports included in this release. Artifact annotations are not available in the current version. Segmentations are limited to the left ventricular wall.

Code and Reproducibility

A reproducible segmentation workflow is available and documented in the companion software capsule. It includes a Dockerfile, automation scripts, environment configuration, and instructions for running nnU-Net on this dataset.

Code repository available at [2,3].

Release Notes

In this database, there is only one version, which is version 1.0.0. However, it contains 103 DICOM images and 101 NIfTI images. This discrepancy arises because two of the exams were not segmented.

Ethics

This database was approved by the appropriate ethics committee for research involving human subjects. The study received ethical approval from the Albert Einstein Institute of Education and Research (IIEP), under CAAE number 65628422.7.0000.0071. All participants signed an Informed Consent Form (ICF), indicating that they were properly informed about the study objectives, procedures, potential risks, and benefits, and voluntarily agreed to participate.

The database was created with the goal of supporting the development of AI-based methods for improving the diagnosis and treatment of coronary artery disease (CAD), enhancing the efficiency and reproducibility of automated segmentation processes, and enabling clinical feasibility studies. By providing access to real-world data, this resource aims to advance the integration of AI in nuclear cardiology, support quality control through improved segmentation compliance, and contribute to artifact correction strategies.

Acknowledgements

The authors acknowledge financial support from the Fundação para a Ciência e a Tecnologia (FCT, Portugal), I.P., under project UIDB/00048/2020 (DOI: 10.54499/UIDB/00048/2020), and from the National Council for Scientific and Technological Development (CNPq, Brazil) through a Research Productivity Fellowship (Grant No. 301644/2022-5). Computational resources were provided by LaMCAD/UFG. The authors also thank the Instituto Israelita de Ensino e Pesquisa Albert Einstein (IIEP, São Paulo, Brazil) for providing access to imaging data and research infrastructure, and the Research Support Foundation of the State of Goiás (FAPEG, Brazil) for funding a master’s scholarship (Grant No. 202310267000733).

Conflicts of Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work related to the construction of the database.

References

1. Dorbala S, Ananthasubramaniam K, Armstrong IS, Chareonthaitawee P, DePuey EG, Einstein AJ, et al. Single photon emission computed tomography (SPECT) myocardial perfusion imaging guidelines: Instrumentation, acquisition, processing, and interpretation. J Nucl Cardiol. 2018;25:1784–1846. doi:10.1007/s12350-018-1284-9
2. Nogueira SA, Luz FAB, Camargo TFO, Oliveira JCS, Campos Neto GC, Carvalhaes FBF, et al. Artificial intelligence applied in identifying left ventricular walls in myocardial perfusion scintigraphy images: Pilot study. PLoS ONE. 2025;20: e0312257. doi:10.1371/journal.pone.0312257
3. Calixto WP, Camargo TFO, Luz FAB, Nogueira SA. Artificial intelligence routines for identifying left ventricular walls in myocardial perfusion scintigraphy. Code Ocean. 2024. doi:10.24433/CO.8499026.v1