MIMIC-IV-ECHO-Ext-LVVOLUMES-A4C-ROI: Annotated Subset of Apical Four-Chamber Echocardiography for PoCUS-Style LV Volume and Function Analysis

@article{PhysioNet-mimic-iv-echo-ext-lvvol-a4c-1.0.0,
  author = {Ekambaram, Kamlin and Arnab, Anurag and Herbst, Philip and Theart, Rensu},
  title = {{MIMIC-IV-ECHO-Ext-LVVOLUMES-A4C-ROI: Annotated Subset of Apical Four-Chamber Echocardiography for PoCUS-Style LV Volume and Function Analysis}},
  journal = {{PhysioNet}},
  year = {2026},
  month = feb,
  note = {Version 1.0.0},
  doi = {10.13026/713s-z339},
  url = {https://doi.org/10.13026/713s-z339}
}

Abstract

MIMIC-IV-ECHO-Ext-LVVOLUMES-A4C-ROIis a curated, credentialed subset of MIMIC-IV-ECHO v0.1 focusing on apical four-chamber (A4C) echocardiographic views with accompanying left-ventricular volumetric labels. The release includes 1,064 de-identified A4C DICOM video sequences from 806 unique patients, each paired with a manually annotated region-of-interest (ROI) mask in both PNG and JSON formats at the original DICOM 4D resolution. ROI-applied, downsampled (256 × 256) MP4 and NPZ derivatives are provided for efficient model development. The full-resolution masks preserve native geometry, enabling regeneration of higher-resolution datasets directly from the source DICOMs.

This resource supports three primary use cases: (i) as an out-of-distribution (OOD) PoCUS external test set for evaluating state-of-the-art LV volume and ejection fraction (EF) models trained on laboratory echocardiography; (ii) training and benchmarking ROI segmentation models for background and overlay removal; and (iii) developing LV-focused video models using preprocessed MP4/NPZ inputs that eliminate DICOM parsing overhead.

Both single-plane (A4C) and biplane (BP) volumetric and EF labels are included, enabling transparent comparison between monoplane interpretability and biplane geometric precision. Access is provided under the PhysioNet Credentialed Health Data License (v1.5.0).

Background

Quantitative echocardiographic assessment of left-ventricular (LV) function remains fundamental for the diagnosis and management of cardiovascular disease [1]. Left-ventricular ejection fraction (LVEF) and volumetric indices derived from apical four-chamber (A4C) and biplane (BP) views provide key diagnostic information for evaluating systolic performance, guiding therapy, and monitoring disease progression. In recent years, deep learning approaches have shown promise for automating these measurements, but their performance often depends heavily on the quality and context of the training data [2]. Models developed on laboratory-grade echocardiograms frequently fail when applied to point-of-care ultrasound (PoCUS) data, where imaging artefacts, operator variability, and device heterogeneity introduce significant domain changes [3].

MIMIC-IV-ECHO-Ext-LVVOLUMES-A4C-ROI addresses this translational gap by providing a quality-controlled and credentialed subset of echocardiographic studies derived from the MIMIC-IV-ECHO v0.1 dataset [4], which in turn is linked to the larger MIMIC-IV Clinical Database [5]. The parent MIMIC-IV-ECHO dataset contains de-identified echocardiographic DICOM studies acquired at Beth Israel Deaconess Medical Centre between 2017 and 2019 using GE Vivid E95, E90, and S7 systems. Each study is matched to patient-level data within MIMIC-IV, enabling multimodal clinical research under PhysioNet’s credentialed access framework [6].

Building upon this foundation, the present release focuses exclusively on apical four-chamber (A4C) videos, the primary view for LV volumetric analysis. It supplements them with manually annotated region-of-interest (ROI) masks that delineate the ultrasound sector. This facilitates both conventional and transformer-based model development for LV function estimation under conditions that better approximate real-world PoCUS imaging. By including single-plane (A4C) and biplane (BP) volumetric labels, the dataset supports investigations into methodological trade-offs between transparency and geometric accuracy.

The following data are available to the research community:

1. an out-of-distribution (OOD) PoCUS test set for evaluating generalisation of models trained on laboratory datasets such as EchoNet-Dynamic [2];
2. a high-quality benchmark for training and evaluating ROI-segmentation networks that isolate diagnostically relevant regions and suppress machine overlays; and
3. a preprocessed resource for developing LV-focused video models, provided as uniformly downsampled MP4 and NPZ sequences that eliminate DICOM parsing overhead.

MIMIC-IV-ECHO-Ext-LVVOLUMES-A4C-ROI, therefore, serves as an open, reproducible benchmark to advance robust and generalisable cardiac video analysis. It provides the research community with a credentialed PoCUS-style reference for evaluating model performance across domains, developing interpretable ROI-segmentation and volume-estimation pipelines, and studying generalisation in real-world imaging conditions. Its manual ROI annotations, dual-plane volumetric labels, and ready-to-train video derivatives make it uniquely positioned to accelerate research in bridging the gap between medical computer vision for laboratory echocardiography and point-of-care ultrasound.

Methods

Data Source and Cohort Selection

Video sequences in this release were drawn from the first three patient folders (p10–p12) of MIMIC-IV-ECHO v0.1 [4]. Each record in MIMIC-IV-ECHO represents a multi-view echocardiography study in DICOM format, containing several video sequences that have been fully de-identified by the MIMIC-IV-ECHO team and redistributed under credentialed access.

For this pilot subset, a single A4C view was manually identified per study—the principal view used for left-ventricular functional assessment. Studies were excluded if they lacked a valid A4C video sequence, did not include left-ventricular volume measurements, or had markedly poor image quality. Of the 1,328 patients from folders p10–p12, the final dataset contains 1,064 studies from 806 unique patients.

This restricted sampling was chosen pragmatically to produce an initial, quality-controlled subset of approximately 1,000 A4C videos for external validation experiments, while establishing the full curation pipeline.

Cardiac Cycle Context

Each DICOM video sequence typically contains about three consecutive cardiac cycles, reflecting the routine echocardiographic acquisition process [1]. In sinus rhythm, this is easily achieved; in arrhythmias such as atrial fibrillation, the operator typically acquires at least three cycles to capture representative motion. The dataset metadata does not indicate which specific cycle the reporting clinician or vendor software used to derive volumetric measurements (EDV, ESV, and LVEF). Therefore, while all frames are preserved, the precise cardiac cycle corresponding to the reported measurements cannot be determined from available metadata. This limitation is inherent to the parent dataset and not introduced by this subset.

Preprocessing and ROI Generation

Each source DICOM study was decoded using pydicom and verified for multi-frame integrity. All frames were retained at the original resolution (1016 × 708 pixels) with photometric interpretation YBR_FULL_422. No resampling, denoising, or temporal interpolation was applied to preserve the integrity of the parent data.

Manual ROI Annotation

For each included study, the ultrasound sector (fan-shaped field of view) was manually segmented using LabelMe [7]. Each ROI was annotated on the first frame only, since the ultrasound sector remains static across the video sequence. The region was drawn as a ten-sided polygon approximating a pie sector, even when the apical boundary appeared truncated in the source DICOM, to ensure consistent radial coverage and to exclude interface elements.

Each study includes both the original LabelMe JSON annotation (masks/json), preserving polygon coordinates and DICOM spatial metadata, and a ready-to-use binary PNG rasterisation (masks/png). Both JSON and PNG masks are at the original DICOM frame resolution (1016 × 708). Each JSON mask follows the standard LabelMe schema (imageHeight, imageWidth, shapes) and is referenced to the original DICOM pixel grid; PNG masks are faithful rasterisations of the same geometry (inside-ROI = 255, outside-ROI = 0).

ROI Application and Export

ROI-applied MP4 and NPZ files were later generated from these static manual masks, then downsampled and centre-cropped to 256 × 256 pixels for LV-focused model training. Pixel intensities were preserved inside the annotated region and set to zero outside to suppress background and interface overlays.

Using the manual masks, we produced ROI-applied sequences in two user-friendly formats:

• MP4 (H.264) for quick visualisation and baselines, and
• NPZ arrays (e.g., frames) for efficient deep-learning pipelines.

No temporal interpolation was applied; the frame rate is taken from CineRate or FrameTime. The FileList.csv provides per-study metadata and volumes (A4C and, where available, BP). All operations used open-source libraries (pydicom, opencv-python, numpy).

Labels and Metadata Extraction

Left-ventricular volumetric measurements were manually extracted from the original DICOM echocardiograms included in MIMIC-IV-ECHO v0.1. For each selected study, the relevant A4C video sequence was reviewed to identify the frames annotated as end-diastole (EDV) and end-systole (ESV) by the reporting clinician or ultrasound system. These annotations, typically visible as on-screen numerical overlays or embedded text metadata, were transcribed into the dataset’s structured FileList.csv file. The ejection fraction (LVEF_A4C) was captured directly from the automated value displayed on the same study, calculated by the system’s internal implementation of the Simpson’s single-plane (method of disks) formula.

Where available, the corresponding biplane (BP) measurements—end-diastolic volume (LVEDV_BP), end-systolic volume (LVESV_BP), and biplane ejection fraction (LVEF_BP)—were extracted from the same DICOM study. These values were typically displayed as part of the embedded report text generated by vendor-specific measurement software, which applies the biplane Simpson’s method using the A4C and apical two-chamber (A2C) video sequences. All extracted values were verified by double review, and any ambiguous or partially rendered overlays were excluded.

Approximately 96% of studies contained both A4C and biplane-derived volumes; the remaining 4% included only single-view A4C measurements due to the absence or poor quality of the A2C view. All volumes are expressed in millilitres (ml) and ejection fractions as percentages (0–100%).

Label Verification

To ensure internal consistency, all recorded left-ventricular ejection fraction (LVEF) values were independently recalculated from the corresponding EDV and ESV values using the standard formula:

$LVEF = \frac{EDV − ESV}{EDV} \times 100\%$

The ejection fraction values included in FileList.csv were recomputed from the extracted end-diastolic and end-systolic volumes using double-precision arithmetic. These values correspond exactly to the percentages displayed on the original echocardiograms but are provided here with full numerical precision rather than being rounded to the nearest integer. Differences between the rounded on-screen EF values and the recomputed EF values were 0.5%. This approach preserves mathematical consistency across all records while allowing users to apply their own custom rounding or discretisation strategies if desired.

Choice of Measurement Method

Users may choose whether to rely on single-plane (A4C) or biplane (BP) volumetric labels when training or evaluating models. While the biplane Simpson’s method is generally regarded as closer to the reference standard for ejection fraction estimation [1], it may incorporate additional assumptions and vendor-specific implementations that are not fully disclosed in the available metadata of the parent dataset. In contrast, the monoplane A4C approach—employed by large public datasets such as EchoNet-Dynamic [2]—offers greater transparency, reproducibility, and alignment with real-world PoCUS constraints where only a single acoustic window may be obtainable. Both sets of measurements are provided to enable comparative studies of accuracy, reproducibility, interpretability, and domain transfer. For reference, the standard Simpson's method formulas for estimating chamber volumes ($V$) for a given frame (EDV or ESV) are as follows:

Monoplane (A4C):

$V_{A4C} = \sum_{i=1}^{N} \frac{\pi}{4} D_i^2 \Delta L$

where $L$ is the long-axis length from apex to mitral annulus, divided into $N$ equal segments of thickness $\Delta L = L/N$, and $D_i$ is the short-axis diameter at slice $i$.

Biplane (A4C + A2C):

$V_{BP} = \sum_{i=1}^{N} \frac{\pi}{4} D_{A4C,i} D_{A2C,i} \Delta L$

which combines orthogonal diameters from both views to reduce geometric assumptions.

Data Description

The dataset comprises the following components:

• Original de-identified A4C DICOM video sequences

• Manual ROI masks (two formats, one per study):

  • Vector (editable): masks/json/<study_id>.json (LabelMe format; original DICOM resolution metadata preserved)

  • Binary (ready-to-use): masks/png/<study_id>.png (8-bit, 0/255; same resolution as the source DICOM)

• ROI-applied MP4 videos (256 × 256 pixels, H.264; derived from the JSON/PNG masks)

• ROI-applied NPZ arrays (one per study, total N = 1,064; each stores [n_frames, 256, 256, 1], where n_frames varies by acquisition)

• Metadata file FileList.csv: per-study identifiers, frame timing, device metadata, and volumetric labels (A4C and, when available, BP)

Note: While the distributed ROI-applied MP4 and NPZ files are downsampled to 256 × 256 for uniformity and computational efficiency, the accompanying manual masks (JSON/PNG) preserve the original DICOM geometry (1016 × 708). Users may reapply these masks to the full-resolution DICOMs to recreate higher-fidelity datasets or adapt them for custom preprocessing pipelines.

File Naming and Traceability

Each record in this dataset is associated with a study_id, which serves as the canonical identifier across all derived files (e.g., MP4, NPZ, PNG, JSON). This identifier corresponds to the parent echocardiographic study in the original MIMIC-IV-ECHO dataset.

For traceability, the metadata include a parent_dicom_path field in FileList.csv, which records the relative path to the corresponding cine file within the parent MIMIC-IV-ECHO directory structure. This field is provided solely to enable linkage for users who already possess the parent dataset; all DICOM files required to reproduce the derived outputs in this release are included locally within the submitted dicom/ directory.

In the parent dataset, echocardiographic cines follow the structure:

files/p12/p12363362/s90004661/90004661_0036.dcm,

where s90004661 denotes the study-level directory and the suffix _0036 identifies a specific cine sequence (e.g., an apical four-chamber view) within that study.

For this curated release, we remove the s prefix and use the resulting numeric identifier (90004661) as the canonical study_id across all derived files (e.g., 90004661.mp4, 90004661.npz, 90004661.json). This design choice reflects that left-ventricular (LV) volumetric measurements in the source metadata are study-level quantities, rather than view-level values—particularly in the case of biplane (BP) measurements, which combine information from multiple projections.

This convention ensures that each record’s volumetric labels (EDV, ESV, EF) remain correctly associated with the originating study, even when multiple cine views are present in the parent DICOM directory.

Directory Structure

Files are organised by type rather than by patient, facilitating incremental updates and efficient programmatic access:

MIMIC-IV-ECHO-Ext-LVVOLUMES-A4C-ROI/
├── docs/
├── dicom/
├── npz/
├── mp4/
├── masks/
│   ├── png/
│   └── json/
├── README.txt
└── FileList.csv

Each study has one file of each type sharing the same base identifier (for example, 90004661.dcm, 90004661.npz, 90004661.png, 90004661.json, and 90004661.mp4).

Data Details

The per-study metadata and measurement fields are stored in FileList.csv. A detailed summary of column definitions is provided in Table 1 and Table 2 below. A visual overview of dataset composition, frame statistics, volume distributions, manufacturer breakdown, and comparison between A4C and biplane measurements is included in docs/dataset_visualisation.png, which contains histograms, scatter plots, Bland–Altman plots, and data-completeness heatmaps.

Technical Validation

Label Consistency

We verified numerical consistency between the provided and recomputed LVEF values across all studies, confirming zero deviation across all studies.

ROI Mask Quality Control

All ROI masks underwent visual quality control by two reviewers and demonstrated complete (100%) concordance for inclusion; no cases required correction. Minor variations at the apical sector edge were within expected tolerance and did not affect field-of-view integrity. Because the ROI-applied MP4 and NPZ outputs are generated directly from the manual PNG masks, ROI fidelity is preserved across all derived formats.

A4C and Biplane Comparison

Agreement between single-plane (A4C) and biplane (BP) volumetric measurements was assessed using Bland–Altman analysis [8]. Summary results are available in Table 3 and docs/dataset_visualisation.png:

These results indicate no systematic bias between the single-plane (A4C) and biplane (BP) measurements. Full scatter plots and Bland–Altman visualisations are available in docs/dataset_visualisation.png, and the corresponding study-level values are provided in FileList.csv to enable independent re-analysis or alternative visualisations by dataset users.

Usage Notes

This dataset supports a range of workflows spanning evaluation, segmentation, and model development:

• Model evaluation: Direct evaluation of LV volume or LVEF estimation models using FileList.csv together with the ROI-applied MP4 or NPZ files.

• ROI segmentation: Training and benchmarking of segmentation models using DICOM frames with corresponding binary PNG masks as ground truth.

• LV-focused video modelling: Development of regression or classification models on ROI-applied NPZ sequences, suitable for transformer or CNN-based architectures.

The MP4 and NPZ formats enable deep learning applications to run without requiring DICOM parsing. Frame rate is derived from either the CineRate or FrameTime DICOM tag, depending on availability. Each video typically contains multiple cardiac cycles, and the exact cycle corresponding to the reported volumetric measurements cannot be identified; therefore, frame-level analyses should be considered semi-supervised.

Beyond direct model evaluation, this dataset provides a versatile substrate for research into domain adaptation, self-supervised cardiac representation learning from a PoCUS distribution, uncertainty-aware prediction, and generative simulation of cardiac motion. Its credentialed linkage to the broader MIMIC-IV-ECHO ecosystem also enables future multimodal extensions combining imaging, physiology, and clinical outcomes.

All metadata is contained within a single CSV file FileList.csv. Below are minimal Python examples illustrating typical usage patterns for metadata parsing, DICOM inspection, ROI masking, and preprocessed NPZ loading.

Each code example below executes without requiring external dependencies beyond pandas, numpy, pydicom, Pillow, matplotlib, and imageio. They enable complete reproducibility of the key steps described in the Methods section, including LVEF verification, DICOM inspection, ROI visualisation, and loading the preprocessed video.

Example A: Parsing FileList.csv

import pandas as pd

# Load metadata
meta = pd.read_csv("FileList.csv")

# Select key fields
cols = [
    "patient_id", "study_date", "study_time", "parent_dicom_path", "study_id",
    "frame_rate", "number_of_frames", "duration_seconds",
    "LVEDV_A4C", "LVESV_A4C", "LVEF_A4C",
    "LVEDV_BP", "LVESV_BP", "LVEF_BP"
]
df = meta[cols].copy()

# Recompute EF and verify precision
df["LVEF_from_volumes"] = (df["LVEDV_A4C"] - df["LVESV_A4C"]) / df["LVEDV_A4C"] * 100
print((df["LVEF_from_volumes"] - df["LVEF_A4C"]).abs().max())

Example B: Load and Display a DICOM Frame

import pydicom
import matplotlib.pyplot as plt

dcm = pydicom.dcmread("dicom/90004661.dcm")
arr = dcm.pixel_array
frame = arr[0] if arr.ndim == 3 else arr

plt.imshow(frame, cmap="gray")
plt.axis("off")
plt.title("First frame from 90004661.dcm")
plt.show()

Example C: Apply ROI Mask to a Frame

import numpy as np
from PIL import Image
import matplotlib.pyplot as plt

frame = arr[0]
mask = np.array(Image.open("masks/90004661.png").convert("L"))
mask_bin = (mask > 0).astype(np.uint8)
masked = frame * mask_bin

fig, ax = plt.subplots(1, 3, figsize=(12, 4))
ax[0].imshow(frame, cmap="gray"); ax[0].set_title("Original"); ax[0].axis("off")
ax[1].imshow(mask_bin, cmap="gray"); ax[1].set_title("ROI Mask"); ax[1].axis("off")
ax[2].imshow(masked, cmap="gray"); ax[2].set_title("Masked"); ax[2].axis("off")
plt.tight_layout()
plt.show()

Example D: Load Processed NPZ Video

import numpy as np
import imageio.v3 as iio

npz = np.load("npz/90004661.npz")
frames = npz["frames"] if "frames" in npz.files else npz[npz.files[0]]
print("NPZ shape:", frames.shape)

# Convert to GIF preview
iio.imwrite("preview.gif", frames, fps=30)

Release Notes

This initial release represents a pilot subset comprising folders p10-p12 of MIMIC-IV-ECHO. While sufficient for proof-of-concept studies, it may not capture the full diversity of the MIMIC cohort. Extracted labels originate from routine operator-dependent clinical measurements rather than centralised reanalysis, which may include inter-observer variability. Frame-level segmentation is not included. An expanded release encompassing the entire cohort with view classification is currently underway and will be available in subsequent release(s).

Ethics

This dataset is derived from the MIMIC-IV-ECHO database and is shared under the same ethical, regulatory, and data use conditions as the parent dataset. All echocardiographic studies were originally collected as part of routine clinical care at Beth Israel Deaconess Medical Center and were subsequently de-identified in accordance with HIPAA Safe Harbor standards.

Access is provided under the PhysioNet Credentialed Health Data License and requires completion of the corresponding training and data use agreement applicable to MIMIC-IV-ECHO.

Acknowledgements

The authors acknowledge the PhysioNet team and the creators of the MIMIC-IV-ECHO database for providing access to the credentialed source material.

Conflicts of Interest

The author(s) have no conflicts of interest to declare.

References

1. R. M. Lang, L. P. Badano, V. Mor-Avi, J. Afilalo, A. Armstrong, L. Ernande, F. A. Flachskampf, E. Foster, S. A. Goldstein, T. Kuznetsova, P. Lancellotti, D. Muraru, M. H. Picard, E. R. Rietzschel, L. Rudski, K. T. Spencer, W. Tsang, and J.-U. Voigt, “Recommendations for cardiac chamber quantification by echocardiography in adults: An update from the American Society of Echocardiography and the European Association of Cardiovascular Imaging,” Journal of the American Society of Echocardiography, vol. 28, no. 1, pp. 1–39.e14, 2015.
2. D. Ouyang, B. He, A. Ghorbani, N. Yuan, J. Ebinger, C. P. Langlotz, P. A. Heidenreich, R. A. Harrington, D. H. Liang, E. A. Ashley, and J. Y. Zou, “Video-based AI for beat-to-beat assessment of cardiac function,” Nature, vol. 580, no. 7802, pp. 252–256, 2020.
3. D. Crockett, C. Kelly, J. Brundage, J. Jones, and P. Ockerse, “A stress test of artificial intelligence: Can deep learning models trained from formal echocardiography accurately interpret point-of-care ultrasound?” Journal of Ultrasound in Medicine, vol. 41, no. 10, pp. 2561–2572, 2022. doi:10.1002/jum.16007.
4. B. Gow, T. Pollard, N. Greenbaum, B. Moody, A. Johnson, E. Herbst, J. W. Waks, P. Eslami, A. Chaudhari, T. Carbonati, S. Berkowitz, R. Mark, and S. Horng, “MIMIC-IV-ECHO: Echocardiogram Matched Subset (version 0.1),” PhysioNet, 2023. doi:10.13026/ef48-v217.
5. A. E. W. Johnson, L. Bulgarelli, T. Pollard, S. Horng, L. A. Celi, and R. G. Mark, “MIMIC-IV (version 1.0),” 2021.
6. A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E. Stanley, “Physiobank, physiotoolkit, and physionet: Components of a new research resource for complex physiologic signals,” Circulation, vol. 101, no. 23, pp. e215–e220, 2000, rRID: SCR_007345.
7. B. C. Russell, A. Torralba, K. P. Murphy, and W. T. Freeman, “LabelMe: A Database and Web-Based Tool for Image Annotation,” International Journal of Computer Vision, vol. 77, no. 1-3, pp. 157–173, 2008. [Online]. Available: https://doi.org/10.1007/s11263-007-0090-8
8. J. M. Bland and D. G. Altman, “Statistical methods for assessing agreement between two methods of clinical measurement,” Lancet, vol. 1, no. 8476, pp. 307–310, 1986.