bgx666/EV9V

Dataset Details

Dataset Description

EV9V (Echocardiographic View 9 Varieties) is a large-scale echocardiographic video dataset for cardiac view classification. It contains 5,138 clinical echocardiographic examination videos comprising 910,579 valid image frames, spanning 9 standard transthoracic echocardiographic views. Videos were acquired during routine clinical examinations at the First Affiliated Hospital of Chengdu Medical College between January 2020 and January 2023, and were performed by licensed ultrasound physicians with varying levels of seniority (from residents to chief physicians). All videos were recorded at a uniform frame rate of 30 frames per second, with lengths ranging from 25 to 3,243 frames (median: 116; mean: 177 ± 190). Extracted frames are stored as JPEG images at 320 × 240 resolution.

The dataset is released to support reproducible research in automated echocardiographic view classification and to provide a comprehensive benchmark covering the primary imaging planes for cardiac assessment.

• Curated by: Bo Gou, Jicheng Zhang, Jianlong Xiong, Tao He, Bentian Liu, Hai Wu, Yijiao Wang, Yu Zhang, Yujia Yang, Yun Dai, Jian Liu, Jie Wang
• Institution: The First Affiliated Hospital of Chengdu Medical College; Sichuan University; West China Hospital of Sichuan University; Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College
• Language(s) (NLP): N/A
• License: cc-by-4.0

Dataset Sources

• Repository: Hugging Face (bgx666/EV9V)
• Paper: Spatio-Temporal Fusion Model for Standard View Classification of Echocardiographic Videos

Uses

Direct Use

The primary use case is multi-class classification of echocardiogram video clips into 9 standard cardiac views. This can assist in:

• Automated view classification during ultrasound acquisition
• Quality assurance in echocardiography workflows
• Training and evaluation of video-based or frame-based medical image classification models
• Benchmarking spatio-temporal architectures for medical video analysis

Out-of-Scope Use

• Not intended for clinical diagnosis without human verification
• Not validated for real-time patient monitoring
• Should not be used as the sole decision-making tool in clinical settings
• Not designed for detection of cardiac abnormalities or disease diagnosis

Dataset Structure

The dataset consists of 9 standard transthoracic echocardiographic views defined in accordance with the 2018 American Society of Echocardiography Guidelines:

Class Code	View Name	Training	Validation	Test
PLHLA	Parasternal Long Axis View	1,120	185	275
PASA	Parasternal Short Axis View	212	16	41
PMPALA	Pulmonary Main Pulmonary Artery Long Axis View	319	46	88
PMVLSA	Parasternal Short Axis View at Mitral Valve Level	262	32	50
PPMLSA	Parasternal Short Axis View at Papillary Muscle Level	183	41	56
PMASA	Parasternal Short Axis View at Apical Level	317	50	72
A4C	Apical 4-Chamber View	858	148	231
A5C	Apical 5-Chamber View	217	29	40
SC4C	Subcostal 4-Chamber View	195	20	35

Each video is stored as a sequence of JPEG frames in Images/{video_name}/.

The dataset is split into three text files, each containing video filenames and their corresponding view labels (one video per line):

• train_labeled.txt — 3,683 videos (504 patients)
• validation_labeled.txt — 567 videos (72 patients)
• test_labeled.txt — 888 videos (127 patients)

Labels are stored alongside video names within the split files in the format {video_name} {label}. To prevent data leakage from correlated videos of the same patient, the dataset was partitioned at the patient level.

Dataset Creation

Curation Rationale

The dataset was curated to address the limited scale and view coverage of existing public echocardiographic datasets, which typically contain only 2–5 views from a few hundred patients. EV9V comprises 5,138 videos and 910,579 frames spanning 9 standard views from 703 patients, making it to our knowledge the largest publicly accessible echocardiographic video dataset for view classification. It further includes the clinically important yet frequently overlooked PMPALA view, implements rigorous three-tier annotation quality control, and captures real-world clinical variability in patient anatomy, acoustic windows, and scanning duration.

Source Data

Data Collection and Processing

Echocardiogram videos were collected during routine clinical examinations at the First Affiliated Hospital of Chengdu Medical College. All examinations were performed by licensed ultrasound physicians following routine clinical echocardiography protocols, encompassing natural variations in patient anatomical structure, acoustic window conditions, and operator proficiency. Frame extraction was performed using FFmpeg, and all extracted frames were resized to a uniform resolution of 320 × 240 pixels. After removing invalid frames, the final dataset contained 910,579 valid image frames.

This study was approved by the Ethics Committee of the First Affiliated Hospital of Chengdu Medical College (Ethics Approval No. 2026CYFYIRB-BA-067).

Who are the source data producers?

Licensed ultrasound physicians (ranging from residents to chief physicians) performed the echocardiogram examinations at the First Affiliated Hospital of Chengdu Medical College under routine clinical protocols. The patient population consists of adults aged 20–68 years undergoing routine cardiac assessment or screening for suspected heart disease.

Annotations

Annotation process

A rigorous three-tier physician quality control system was implemented to ensure annotation accuracy and consistency. Videos were initially annotated independently in a double-blind manner by multiple licensed attending physicians with clinical experience in echocardiography, using the Anvil video annotation research tool. Any discrepancies or controversies were reviewed and adjudicated by higher-ranking associate chief physicians or senior experts. Finally, a secondary verification was conducted by senior experts in cardiac ultrasound diagnosis to confirm the consistency and clinical reliability of all labels.

Who are the annotators?

Licensed attending physicians with clinical experience in echocardiography performed the initial double-blind annotation. Discrepancies were adjudicated by associate chief physicians or senior experts. Final verification was conducted by senior experts in cardiac ultrasound diagnosis.

Personal and Sensitive Information

To protect patient privacy, complete anonymization and removal of all patient-identifiable information were performed prior to any data processing and annotation work. Video filenames are anonymized with no patient identifiers. No demographic, clinical history, or diagnostic information is included. The dataset contains only echocardiogram images and view labels.

Bias, Risks, and Limitations

• The dataset was collected from a single institution, which may introduce bias in acquisition protocols, patient demographics, and equipment characteristics.
• Class distribution is imbalanced, reflecting real-world clinical acquisition patterns: PLHLA and A4C are the most prevalent views, while PASA, PMPALA, and SC4C are comparatively underrepresented.
• Video quality varies due to differences in patient anatomy, acoustic window conditions, sonographer experience, and acquisition conditions.
• The dataset does not include pathological findings or diagnostic outcomes—only view classification labels.
• Videos were acquired following routine clinical practice rather than strict standardized scanning protocols, reflecting real-world variability but introducing additional noise.

Recommendations

Users should be aware of the class imbalance and consider appropriate re-sampling or weighting strategies. Models trained on this dataset should be validated on external multi-institutional data before clinical deployment. Performance may vary across different patient populations and ultrasound systems.

Citation

BibTeX:

@article{he2026ev9v,
  author = {Gou, Bo and Zhang, Jicheng and Xiong, Jianlong and He, Tao and Liu, Bentian and Wu, Hai and Wang, Yijiao and Zhang, Yu and Yang, Yujia and Dai, Yun and Liu, Jian and Wang, Jie},
  title = {Spatio-Temporal Fusion Model for Standard View Classification of Echocardiographic Videos},
  year = {2026}
}

Dataset Card Authors

Bo Gou, Jicheng Zhang, Jianlong Xiong, Tao He, Bentian Liu, Hai Wu, Yijiao Wang, Yu Zhang, Yujia Yang, Yun Dai, Jian Liu, Jie Wang

Dataset Card Contact

Tao He (Sichuan University)