LGE-Mi40: Myocardial Infarction Dataset
Description
This dataset provides late gadolinium enhancement cardiac magnetic resonance (LGE-CMR) images for the study of myocardial infarction (MI) analysis, supporting the development and evaluation of segmentation, intensity-based representation, and image enhancement methods. The data consist of 2D short-axis LGE-CMR images acquired retrospectively from 40 patients using a 3.0T MRI system at Pusat Kanser Tunku Abdullah Ahmad Badawi (formerly AMDI), Universiti Sains Malaysia where the examinations were performed between 2018 and 2024. Images were retrieved from the institutional PACS using OsiriX (v12.5.2) and later were de-identified. Acquisition employed a breath-hold technique with SENSE acceleration, with slice thickness and inter-slice spacing both set to 8 mm. All of the image have native resolution of 320 x 320. Inclusion criteria comprised patients aged ≥18 years with post-infarction LGE acquired 1–3 months after the event, showing scar consistent with chronic fibrosis. Exclusion criteria included severe imaging artifacts, congenital heart disease, and significantly distorted cardiac anatomy due to prior thoracic surgery. Three reconstruction types are included per anatomical slice: magnetization-prepared inversion recovery (MR-IR), conventional inversion recovery (CR-IR), and modified fast field echo (M-FFE). MR-IR and CR-IR provide complementary inversion recovery contrast distributions, while M-FFE provides a bright-contrast reconstruction that does not apply null-point suppression, offering clearer left ventricular boundary delineation, enabling multi-modal input strategies for learning-based methods. Three primary data components are provided: (1) raw LGE-CMR images without windowing, allowing custom preprocessing; (2) manually delineated and radiologist-verified multi-class segmentation masks with four RGB-encoded classes (myocardium, cavity, infarct scar, background); and (3) intensity-encoded ground truth (IEGT) images with enhanced scar conspicuity, generated via contrast normalization, UNLM denoising, CDF-based contrast refinement, and mask-guided intensity encoding. The dataset is provided in two format variants: 16-bit TIFF (raw-tif) preserving full DICOM-derived intensity values, and 8-bit PNG (raw, iegt, mscgt) for direct use in learning-based pipelines. Additional files include slice-level DICOM metadata (.csv) and patient-level radiological reports (.pdf). Each image follow the naming convention 'Pxx_aa_bb', where 'aa' denotes reconstruction type (01: MR-IR, 02: CR-IR, 03: M-FFE), enabling direct pairing across all components.
Files
Steps to reproduce
The paper for this dataset, original source code and supplementary materials detailing the preprocessing steps in showing reproducibility are provided via GitHub repository link.
Institutions
- Universiti Teknologi MARASelangor, Shah Alam
- Universiti Sains MalaysiaPenang, George Town