Database Open Access

Point-of-Care Gastric Ultrasound in Healthy Volunteers: Cine Loops and Still Images Across Four Controlled Gastric Content States

, 

Published Nov. 17, 2026 · Version 1.0.0
When using this resource, please cite:

Singleton, B., & Wall, T. (2026). Point-of-Care Gastric Ultrasound in Healthy Volunteers: Cine Loops and Still Images Across Four Controlled Gastric Content States (version 1.0.0). PhysioNet. RRID:SCR_007345. https://doi.org/10.13026/xpj0-mv11

Please include the standard citation for PhysioNet:

Pollard, T., Moody, B. E., Lehman, L., Gow, B., Fernandes, C., Xie, C., Johnson, A., Mark, R. G., & Heldt, T. (2026). PhysioNet as a global platform for biomedical research. Nature Health. https://doi.org/10.1038/s44360-026-00096-z. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC13641116/

Abstract

Point-of-care gastric ultrasound is used before anaesthesia to judge whether the stomach is empty, holds clear fluid, or holds solid content, the last being the finding that most changes airway management. Work on automated interpretation has been limited by a lack of openly available data.

This dataset contains 720 gastric ultrasound acquisitions from 30 fasted healthy adult volunteers. Each volunteer was imaged in four gastric states — fasted, after 250 mL of water, early after a solid meal, and late after that meal — in both the supine and right lateral decubitus positions. Each of the eight state-position combinations yielded one still image and two five-second cine loops, giving 240 stills and 480 loops. All imaging was performed by one operator on a single curvilinear system.

Gastric content was set by the study protocol rather than judged from the image, so every acquisition carries a protocol-derived state label. Cine loops are released in full. Every frame is cropped to the acoustic sector, which removes the acquisition date and time, the session code, the manufacturer and hospital logos, the depth scale, and the operator's on-screen annotation. Machine settings and the timing of each acquisition relative to ingestion are provided as tabular files; participant demographics are reported in aggregate only.

A future version will add gastric state labels assigned independently by experienced operators reading the images blind.


Background

Pulmonary aspiration of gastric contents is the most common cause of anaesthesia-related airway death [@cook2011]. Fasting guidelines manage the risk at the level of populations but say little about an individual patient, and residual gastric content persists in a substantial minority of patients taking glucagon-like peptide-1 receptor agonists despite guideline-adherent fasting [@sen2024]. Point-of-care gastric ultrasound addresses this directly, by imaging the antrum and classifying its contents as empty, clear fluid, or solid [@perlas2009, @perlas2016]. Qualitative classification precedes any quantitative assessment: the validated relationship between antral cross-sectional area and gastric volume holds only for clear fluid, and where solid content is present the volume is not the relevant question.

The technique is nevertheless slow to spread, largely because it is operator dependent; roughly 33 supervised examinations are needed before an anaesthesiologist reaches reliable qualitative performance [@arzola2013]. Automated interpretation is an obvious response, and has been applied successfully in cardiac, lung, obstetric and regional anaesthesia ultrasound. Published work in gastric ultrasound addresses segmentation of the antrum and analysis of antral motility [@zhang2025, @zou2025]; the one framework that attempts content assessment restricts itself to measured volumes of ingested water and does not address solid content [@xiao2025].

Openly available data is a practical obstacle. A 2025 survey identified 72 public ultrasound datasets, spanning cardiac, abdominal, thyroid, obstetric, lung, breast, musculoskeletal and neurological imaging, and none of the stomach [@alsharid2025]. This dataset was collected to make gastric ultrasound data available. It is small and comes from a single centre, single operator and single machine, so it is a starting point rather than a basis for validated models.


Methods

Study design and setting

Prospective, single-centre observational study at Mater Misericordiae University Hospital, Dublin, Ireland, conducted between September 2025 and March 2026. Each participant served as their own control, undergoing serial gastric ultrasound across four gastric states within a single session. All examinations were performed by one operator.

Participants

Adult volunteers aged 18 years or over were recruited by convenience sampling. Inclusion required the ability to give informed consent and to tolerate both supine and right lateral decubitus positioning. Exclusion criteria were pregnancy, previous gastrointestinal surgery, known gastric motility disorder including diabetes mellitus and gastroparesis, current use of a medication that delays gastric emptying (glucagon-like peptide-1 receptor agonist, opioid, anticholinergic), and body mass index of 35 kg/m² or above. Participants fasted from both solids and fluids for at least eight hours before the session. Thirty volunteers were recruited.

Equipment and settings

All imaging used an Esaote MyLab A70 with a C 1-8A curvilinear transducer, on the general abdominal preset in B-mode with tissue enhancement imaging engaged, at an imaging depth of 166 mm and 100% acoustic power. Frames were exported at 1200 × 932 pixels. Cine loops were captured at a nominal 31 frames per second; the machine's frame-rate readout was 32 Hz, and the exported loops carry a frame rate of 30, 31 or 32 frames per second, given per file in acquisitions.csv.

Gain was set once per session and not adjusted during scanning: automatic gain was engaged for 14 participants and the preset gain left in place for 15, with one participant switching from preset to automatic after the fasted state. Gain mode is balanced across the four states and both positions. Full settings are in acquisition_settings.txt, with per-file values in acquisitions.csv.

Scanning protocol

The gastric antrum was identified using the left lobe of the liver, the aorta and the superior mesenteric artery as landmarks [@perlas2009]. Each participant was imaged in four sequential gastric states, in each state in both the supine and the right lateral decubitus position, with three acquisitions per state-position combination: one still image and two five-second cine loops. This gives 24 files per participant and 720 in total.

  • S1, fasted (FAS). After a confirmed fast of at least eight hours from solids and fluids. A fasted stomach is not necessarily an empty one: baseline gastric secretion of up to 1.5 mL/kg is physiologically normal.
  • S2, clear fluid (CLR). After ingestion of 250 mL of water. The volume was standardised: every participant received the same 250 mL. Target window 15–20 minutes, acceptable window 10–30 minutes.
  • S3, solid, early phase (SOLE). After a solid meal of the participant's own choosing. The sole requirement was that the meal involve mastication, so that air would be swallowed with it: swallowed air rises in the antrum and produces the hyperechoic band and posterior shadowing that characterise the early post-prandial stomach [@perlas2009]. Requiring chewing is therefore the minimum constraint that reliably produces the early-phase appearance, while leaving the meal itself unconstrained. Composition and volume were neither prescribed nor recorded, only the time at which the meal was completed. Target window 10–15 minutes, acceptable window 8–25 minutes after completion.
  • S4, solid, late phase (SOLL). 70–90 minutes after meal completion, acceptable window 60–120 minutes, once swallowed air has cleared. Because meals were not standardised, the appearance of solid content at a given interval varies between participants.

Achieved intervals for every participant are in subjects.csv, with whether each fell in the target or the wider acceptable window.

Labels

Gastric content was controlled by the investigator, so each acquisition is labelled with the protocol state that applied when it was taken. This is a record of what the participant had eaten or drunk and when, and it is the dataset's ground truth in version 1.0.0. It is not a reading of the image, and what a given antrum looks like at a given interval after ingestion varies between people; see Usage Notes.

De-identification and image preparation

No patient identifier was entered on the ultrasound system. A randomly generated six-character session code was used in place of participant demographics at the point of acquisition, and no key linking session codes to participants was created or retained.

Every released frame is cropped to the acoustic sector. On this machine everything outside it is annotation: the manufacturer and hospital logos, the acquisition date and clock time, the session code, the imaging settings, the depth scale, the frame-rate readout, and the operator's on-screen note of gastric state and body position. Cropping therefore removes both the identifying elements and the on-screen text naming the gastric state and body position. The overlay of all 720 machine exports was reviewed beforehand to confirm that no free-text name had been entered in any session; none had.

Output is 934 × 766, the sector bounding box plus a small border, with pixels outside the sector set to black. Stills are lossless PNG in colour. Cine loops were re-encoded once, direct from the machine export, at H.264 constant rate factor 14, preserving frame count, frame rate and dimensions; container metadata was discarded, which also removes the MP4 creation_time tag. code/build_release.py is the script that was run, and code/sector_mask.png the mask it applied. The machine exports themselves are not released, so the script documents the procedure rather than allowing it to be re-run.


Data Description

Directory layout

README.md
subjects.csv                      one row per participant (30)
acquisitions.csv                  one row per file (720)
acquisition_settings.txt          machine settings, which the crop removes from the images
LICENSE.txt                       added by PhysioNet at publication
SHA256SUMS.txt                    added by PhysioNet at publication
images/
  <subject>/
    <subject>_s<state>_p<position>_i1.png     still, 934x766, lossless
    <subject>_s<state>_p<position>_i2.mp4     cine loop, 934x766, ~5 s at 30-32 fps
    <subject>_s<state>_p<position>_i3.mp4     cine loop
code/
  build_release.py                produced images/ from the machine exports
  sector_mask.png                 sector mask, 1200x932 export coordinates
  sector_mask_cropped.png         the same mask, matching the released frames
  requirements.txt

File naming

<subject>_s<state>_p<position>_i<acquisition>, all lowercase

Field Values
subject six-character session code, 30 distinct values
s1 s2 s3 s4 fasted, clear fluid, solid early, solid late
p1 p2 supine, right lateral decubitus
i1 i2 i3 still image, first cine loop, second cine loop

The same codes appear in uppercase (S1, P1, I1) in the state, position and acquisition columns of acquisitions.csv.

Counts

Participants 30 (18 male, 12 female)
Age mean 33.4 years, SD 4.3, range 25–47
Age and sex aggregate only; not given per participant
Still images 240 (60 per state, 8 per participant)
Cine loops 480 (120 per state, 16 per participant)
Total cine frames 77,355
Total cine duration 41 minutes
Frame size 934 × 766
Released size 1.3 GB

Every participant has a complete set, with no missing files.

subjects.csv

Column Description
subject_id six-character session code, matches the images/ directory name
fasted_hours_min minimum confirmed fast before S1, hours (8 for all participants)
clear_fluid_volume_ml volume of water ingested before S2; standardised at 250 mL for every participant
min_fluid_to_clr minutes from ingestion of water to the S2 scan
min_meal_to_sole minutes from meal completion to the S3 scan
min_meal_to_soll minutes from meal completion to the S4 scan
clr_window, sole_window, soll_window target or acceptable

acquisitions.csv

One row per file, with file_path, subject_id, state, state_code (FAS/CLR/SOLE/SOLL), state_label, position, position_label, acquisition, modality, width, height, n_frames, frame_rate_hz, duration_s, gain_mode, mechanical_index, thermal_index_soft_tissue, sha256 (the file's SHA-256 checksum) and provenance_note.

One still image, 2r68rd_s2_p1_i1, is the middle frame of its paired cine loop rather than a native machine still, the original having been lost before archiving. It is the only file to which this applies and it is noted in provenance_note.


Usage Notes

Intended use cases

The dataset is intended to support research and education in point-of-care gastric ultrasound. Anticipated uses include:

  • Development and validation of machine-learning models that classify gastric content from still images or cine loops. Because every participant was imaged in all four protocol-controlled states (S1 fasted, S2 clear fluid, S3 early solid, S4 late solid), within-participant comparisons are possible and each acquisition carries a known experimental condition. As set out below, that condition is not an expert reading of the image, and data splits should be made by subject_id.
  • Comparison of still-image and cine-loop inputs. Paired stills and cine loops from the same examination allow users to test whether temporal information improves classification.
  • Effect of patient position. Each state was imaged in both supine and right lateral decubitus positions, supporting study of how position affects the appearance of gastric content and model performance.
  • Benchmarking and transfer learning for small-sample medical ultrasound tasks, including pre-training before fine-tuning on clinical datasets.
  • Education and training. The images form a reference library of gastric appearances under known conditions for clinicians learning gastric ultrasound.
  • Studies of human interpretation, such as inter-rater agreement and comparison of clinician and model performance. A planned version 1.1.0 will add blinded appearance-based labels from independent expert raters.

The dataset is not intended to support clinical decision-making on its own. Given the limitations below, models developed on it should be externally validated in clinical populations before any clinical use.

Repeated measures within participants

The design is repeated-measures: every participant contributes 24 acquisitions across the four states and two positions, and consecutive frames within a cine loop are near-identical. Analyses that treat frames or acquisitions as independent will be optimistic. Group by subject_id.

What the labels represent

The state label records what the participant had ingested and when. It is not a statement about what is visible in a particular image, and the two can diverge. A fasted stomach may hold a small volume of normal gastric secretion, and a modest meal may have substantially emptied by 70–90 minutes in a healthy young adult, so some S4 acquisitions will show little solid content. Users comparing methods should treat the state label as the experimental condition, not as an expert reading.

Limitations

  • Single centre, single operator, single machine and preset. Generalisation across operators, devices and institutions is untested.
  • Healthy young volunteers with a body mass index below 35 kg/m², no relevant comorbidity, no medication delaying gastric emptying and no pregnancy. The populations in which gastric ultrasound is most useful — emergency presentations, delayed gastric emptying, glucagon-like peptide-1 receptor agonist use, higher body mass index — are absent.
  • Content type only. No gastric volume measurement, antral cross-sectional area measurement or segmentation of the antrum is provided.
  • The clear-fluid challenge was standardised at 250 mL of water, but the solid meal was not: participants ate a meal of their own choosing, and its composition and volume were neither prescribed nor recorded. The intervals in subjects.csv state how long after eating each scan was taken, and are what define the early and late solid states, but without knowing what was eaten they cannot support estimates of gastric emptying rate.
  • Cine loops have been re-encoded once, as described under Methods.

Release Notes

1.0.0 — first release. All 720 acquisitions from 30 participants, with protocol-derived state labels, acquisition timings, per-file machine settings, the sector mask and the code that produced the images.

Planned for 1.1.0. Gastric state labels assigned to the still images independently by experienced gastric ultrasound operators, reading the images blind and without knowledge of the protocol condition. Adding a second, appearance-based set of labels alongside the protocol states supports three things this release cannot: measuring how far expert reading of an image agrees with the controlled gastric state, measuring inter-rater reliability between operators on the same images, and giving a second reference against which training approaches can be compared. The operators who provide those labels will be authors of that version.


Ethics

The study was approved by the Research Ethics Committee of Mater Misericordiae University Hospital, Dublin, Ireland (reference 1/378/2513, approval granted 25 August 2025). It was conducted in accordance with the Declaration of Helsinki and the Irish Health Research Regulations 2018.

All participants gave written informed consent before enrolment. The consent form included specific statements that the participant's anonymised ultrasound images might be used to train, test and validate a computer model, that the model's learning from those images could not later be removed, and that "the anonymised dataset and trained model may be shared publicly via a secure and ethically approved platform to support future scientific research and development". The participant information leaflet named PhysioNet as the platform on which the images would be shared.

No medical records were accessed and no identifiable health data were collected. Images were acquired against a randomly generated session code rather than participant demographics, and no key linking session codes to participants was created or retained, so the imaging dataset was anonymised from the point of acquisition. Names, contact details and signed consent forms are held separately and securely by the study team and form no part of this release. As described under Methods, the burned-in acquisition date and time have been removed from every released frame. The released tables contain no dates and no per-participant demographics: age and sex are reported in aggregate only, as the study documents provided for, and the tables hold nothing beyond the protocol state, body position, acquisition timing and machine settings of each session code.

The hospital Data Protection Officer was consulted and confirmed that this release may proceed on the basis that the data is anonymous.

The authors declare that this dataset contains no protected health information.


Acknowledgements

The authors thank the volunteers who gave their time to this study, many of whom were anaesthesia trainees and non-consultant hospital doctors who fasted before, and ate during, a working day.

This work was supported by University College Dublin AI Healthcare Network Seed Funding (reference 90954). The funder had no role in the design or conduct of the study, in the collection or analysis of the data, or in the decision to publish this dataset.


Conflicts of Interest

The authors have no conflicts of interest to declare.

No commercial entity was involved in the funding, design, conduct or analysis of this study. The investigators have not sought a patent on this work and are not selling the dataset or any model derived from it. Esaote is named only as the manufacturer of the ultrasound system used and had no role in the work.


References

  1. @cook2011 Cook, T. M., Woodall, N., & Frerk, C. (2011). Major complications of airway management in the UK: results of the Fourth National Audit Project of the Royal College of Anaesthetists and the Difficult Airway Society. Part 1: Anaesthesia. British Journal of Anaesthesia, 106(5), 617–631. https://doi.org/10.1093/bja/aer058
  2. @sen2024 Sen, S., Potnuru, P. P., Hernandez, N., Goehl, C., Praestholm, C., Sridhar, S., & Nwokolo, O. O. (2024). Glucagon-Like Peptide-1 Receptor Agonist Use and Residual Gastric Content Before Anesthesia. JAMA Surgery, 159(6), 660. https://doi.org/10.1001/jamasurg.2024.0111
  3. @perlas2009 Perlas, A., Chan, Vincent W. S., Lupu, Catalin M., Mitsakakis, N., & Hanbidge, A. (2009). Ultrasound Assessment of Gastric Content and Volume. Anesthesiology, 111(1), 82–89. https://doi.org/10.1097/ALN.0b013e3181a97250
  4. @perlas2016 Perlas, A., Van de Putte, P., Van Houwe, P., & Chan, V. W. S. (2016). I-AIM framework for point-of-care gastric ultrasound. British Journal of Anaesthesia, 116(1), 7–11. https://doi.org/10.1093/bja/aev113
  5. @arzola2013 Arzola, C., Carvalho, J. C., Cubillos, J., Ye, X. Y., & Perlas, A. (2013). Anesthesiologists' learning curves for bedside qualitative ultrasound assessment of gastric content: a cohort study. Canadian journal of anaesthesia = Journal canadien d'anesthesie, 60(8), 771–779. https://doi.org/10.1007/s12630-013-9974-y
  6. @zhang2025 Zhang, H., Zhou, T., Wang, S., Liu, D., Sun, X., Yan, Y., Meng, L., Fan, C., Ni, Z., & Tian, J. (2025). SATU-net: a shadow adaptive tracing U-net for gastric cavity segmentation based on the principle of ultrasound imaging. Scientific Reports, 15(1). https://doi.org/10.1038/s41598-025-20687-2
  7. @zou2025 Zou, T., He, H., Yang, J., Wu, Y., Lv, C., Zhao, L., & Yin, W. (2025). Artificial intelligence real-time automated recognition of the gastric antrum cross-sectional area and motility rhythm via bedside ultrasound: a pilot study. Scientific Reports, 15(1). https://doi.org/10.1038/s41598-025-98974-1
  8. @xiao2025 Xiao, N.-F., Huang, D.-X., Wang, L.-T., Gui, M.-J., Fu, Q., Xie, X.-L., Liu, S.-Q., Wang, S., Hou, Z.-G., Wang, Y.-W., & Zhou, X.-H. (2025). REASON: Probability map-guided dual-branch fusion framework for gastric content assessment (Version 1). arXiv. https://doi.org/10.48550/arXiv.2511.01302
  9. @alsharid2025 Alsharid, M., Guo, X., Men, Q., Saha, P., Mishra, D., Ahuja, R., Ouyang, C., & Noble, J. A. (2025). On the public dissemination and open sourcing of ultrasound resources, datasets and deep learning models. Npj Digital Medicine, 8(1). https://doi.org/10.1038/s41746-025-02162-4

Files

Total uncompressed size: 1.2 GB.

Access the files