Ultrasound - Academic Torrents

Edus2 Ultrasounds (Dataset)

Ultrasound Videos Database: Collection of 32 medical ultrasound video files for simulations, case discussions, and training. Includes cardiac normal, tamponade, FAST exams (RUQ free fluid), AAA, and Edus2 open-source set. Free for non-commercial educational use. This license applies to all video in this directory. Copyright 2011,2012 Paul Kulyk and Paul Olszynski All videos made available under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License.

IUGC: A benchmark of landmark detection in end-to-end intrapartum ultrasound biometry (Dataset)

In 2018, the World Health Organization (WHO) published 56 recommendations to improve the quality of intrapartum care and enhance women’s childbirth experiences. In response, the WHO developed the Labour Care Guide (LCG) in 2020, a next-generation tool designed to promote evidence-based, respectful, and woman-centered care during labor and delivery. The LCG was created through expert consultations, primary research with maternity healthcare providers, and usability studies across multiple countries. It serves as a practical tool for monitoring labor progress and maternal and fetal well-being by recording key clinical parameters. When deviations from normal labor progression are detected, the LCG highlights these issues, prompting timely interventions to ensure safe and effective care. Intrapartum ultrasound for labor progression analysis is a crucial examination in labor management. The core operation in this analysis is the identification of landmarks from intrapartum ultrasound images. These landmarks serve as the basis for subsequent qualitative evaluations of angles and distances, which offer valuable diagnostic information regarding labor arrest and influence decisions about the timing and type of intervention. However, obtaining reliable landmark annotations typically demands experienced physicians, and even for proficient obstetricians, manual landmark identification is a time-consuming and labor-intensive endeavor. Consequently, the development of fully automatic and precise landmark localization techniques has been an area of significant and persistent need. The Intrapartum Ultrasound Grand Challenge (IUGC) 2025 is a collaborative initiative involving the "Deep Learning in Intrapartum Ultrasound Image Analysis" cooperative group and prominent clinical societies such as the International Society of Ultrasound in Obstetrics & Gynecology (ISUOG), the World Association of Perinatal Medicine (WAPM), the Perinatal Medicine Foundation (PMF), and the National Institute for Health and Care-Excellence (NICE). The objective of this partnership is to formulate and promote clinically relevant challenges, thereby maximizing the potential clinical impact of innovative algorithmic contributions from participating teams. Since its inception at MICCAI 2023, the IUGC has advanced the Pubic Symphysis - Fetal Head Segmentation (PSFHS) by facilitating and benchmarking algorithmic progress and providing high-quality annotated image datasets. In MICCAI 2024, the IUGC expanded to incorporate multiple benchmarks: (1) The analysis objects were extended from images to videos; (2) The tasks were augmented from image segmentation to classification, segmentation, and biometric parameter measurement; (3) The quantitative parameters were increased from one (i.e., Angle of Progression (AOP)) to two (i.e., AOP and head - symphysis distance (HSD)); and (4) The data sources were broadened from being solely from Asia to include Asia, Europe, and Africa. This novel and inventive design has established a benchmarking ecosystem for the systematic comparison of algorithms across diverse tasks and clinical challenges. The significance of the IUGC 2025 lies in its concentration on addressing the actual clinical assessment of labor progress, covering (1) end-to-end measurements (which are currently indirect measurements based on segmentation results); (2) all fetal descent stations during the childbirth process (comprising five “minus”, one “zero", and three “plus” stations for reliable longitudinal assessment of labor progress); (3) computational tasks (such as regression, detection); and (4) learning methods (semi-supervised, weakly-supervised, and barely-supervised learning). In line with the IUGC s goal of addressing clinical requirements, authoritative and leading clinical organizations have allied with the IUGC. We have extended the IUGC 2024 Challenge from an indirect ultrasound measurement based on segmentation results to an end-to-end measurement based on landmarks. Specifically, we provide 300 labeled cases and 31,421 unlabeled cases in the training set, 100 visible cases for validation, and 501 hidden cases for testing. The targets are the coordinates of three landmarks and the corresponding biometric parameter. In addition to the typical Mean Radial Error (MRE) and the absolute difference between predicted and manually measured parameters, our evaluation metrics also emphasize inference speed. In summary, the IUGC 2025 challenge exhibits three primary characteristics: (1) Task: Employing semi-supervised landmark detection. (2) Dataset: Curating a large-scale and diverse fetal ultrasound dataset that accounts for all fetal descent stations during the childbirth process. It comprises 28,919 ultrasound images from over 20 medical groups. (3) Evaluation measures: Focusing on detection accuracy. (4) Multiple raters independently annotate a subset of test cases to compare algorithmic performance against human expert inter-rater variability. Bai J, Tang Y, Liu X, Hu J, Li Y, Chen X, Wang Y, Ma C, Li Y, Guo B, Jiao J, Huang Y, Wang K, Li L, Ma Y, Han X, Shao H, Yang Z, Liu Q, Hu Y, Kuang J, Song S, Krishna A, Khan ZA, Li Z, Zhang Z, Zhang H, Cheng Y, Zhang X, Chen X, Yan H, Tong L, Du B, Deng B, Chen Y, Peng Z, Rezaei S, Gan J, Cai W, Wang F, Curran KM, Silvestre G, Khobo I, Lu Y, Ni D, Huang Y, Yaqub M, Ma J, Lekadir K, Li S. IUGC: A benchmark of landmark detection in end-to-end intrapartum ultrasound biometry. Med Image Anal. 2026 May;110:103960. doi: 10.1016/j.media.2026.103960. Epub 2026 Jan 23. PMID: 41604894.

Annotated Ultrasound Liver images (Dataset)

We public the ultrasound liver images, which were annotated to show the outlines, livers, and liver mass regions. Xu Yiming, Zheng Bowen, Liu Xiaohong, Wu Tao, Ju Jinxiu, Wang Shijie, Lian Yufan, Zhang Hongjun, Liang Tong, Sang Ye, Jiang Rui, Wang Guangyu, Ren Jie, & Chen Ting. (2022). Annotated Ultrasound Liver images [Data set]. Zenodo.

TRUSTED: The Paired 3D Ultrasound and CT Human Data for Kidney Segmentation and Registration Research (Dataset)

We propose TRUSTED (the Tridimensional Renal Ultra Sound Tomod Ensitometrie Dataset), comprising paired transabdominal 3DUS and CT kidney images from 48 human patients (96 kidneys), including segmentation, and anatomical landmark annotations by two experienced radiographers. Abstract Inter-modal image registration (IMIR) and image segmentation with abdominal Ultrasound (US) data have many important clinical applications, including image-guided surgery, automatic organ measurement, and robotic navigation. However, research is severely limited by the lack of public datasets. We propose TRUSTED (the Tridimensional Renal Ultra Sound TomodEnsitometrie Dataset), comprising paired transabdominal 3DUS and CT kidney images from 48 human patients (96 kidneys), including segmentation, and anatomical landmark annotations by two experienced radiographers. Inter-rater segmentation agreement was over 93% (Dice score), and gold-standard segmentations were generated using the STAPLE algorithm. Seven anatomical landmarks were annotated, for IMIR systems development and evaluation. To validate the dataset’s utility, 4 competitive Deep-Learning models for kidney segmentation were benchmarked, yielding average DICE scores from 79.63% to 90.09% for CT, and 70.51% to 80.70% for US images. Four IMIR methods were benchmarked, and Coherent Point Drift performed best with an average Target Registration Error of 4.47 mm and Dice score of 84.10%. The TRUSTED dataset may be used freely to develop and validate segmentation and IMIR methods. Ndzimbong, W., Fourniol, C., Themyr, L. et al. TRUSTED: The Paired 3D Transabdominal Ultrasound and CT Human Data for Kidney Segmentation and Registration Research. Sci Data 12, 615 (2025).

Abdominal Ultrasound Image Dataset for Organ Classification and Disease Detection (Dataset)

This is a dataset of Ultrasound (US) images of abdominal organs. US imaging is widely accessible and a very common diagnostic tool, as it is non-invasive and does not involve radiation risk. This dataset was curated solely for research in deep learning, with potential applications in supervised, semi-supervised, and unsupervised learning to support disease detection in resource-constrained settings. The dataset comprises 5,005 images of different abdominal organs, namely: Abdominal Aorta (0), Gallbladder (1), Hepatic Vein (2), Kidneys (3), Liver (4), Ovaries (5), Pancreas (6), Portal Vein (7), Spleen (8), and the Urinary System (9), which includes the Urinary Bladder, Prostate, and Uterus. Images were collected from 563 patients at MH Samorita Medical College and Hospital and in Dhaka, Bangladesh. The author tried her best to curate this dataset systematically and organize uniquely. Every folder and subfolder has correctly numbered, serially ordered images, making it the first dataset one of its kind in terms of structure and usability. This ensures reproducibility and reliability for researchers worldwide In total, the dataset is organized into five distinct formats/folders, described below. Two radiologists were examining the patients. ## Radiologist one: - organ_classification_1: Contains 2,784 images of the 10 organs listed above. Designed for classification tasks. - anomaly_detection_1: Contains two sub-folders: normal (2,014 images) and abnormal (799 images). Designed for anomaly detection tasks. - organ_classification+anomaly_detection: Contains 2 sub-folders. One represents the normal organs (1,948 images) and one represents abnormal cases (981 images) including a newly added ascites folder. Ascites is a condition where the abdominal cavity becomes overly filled with fluid, and it was separated as a distinct abnormal class. This hybrid dataset (organ_classification+anomaly_detection) is an experimental extension combining both tasks. While curated carefully, users are advised to double-check labels for their specific tasks. ## Radiologist two: - anomaly_detection_2: Contains two sub-folders: normal (656 images) and abnormal (269 images). This batch was collected last and was used for semi-supervised anomaly detection tasks. - organ_classification_2: Contains 10 sub-folders representing the 10 different organs, with a total of 1,293 images. - Patient_Wise: This folder contains 170 patient images, their diagnosis as metadata in an xlsx file and a text file, Update Version 01_USG.txt. ## Acknowledgements: This dataset was developed as part of the author s research under the supervision of Dr. John E. Ball, Mississippi State University. The author is grateful for his guidance, encouragement, and support throughout the course of this work. The author would like to thank logistical support of her father, Dr. Md. Enayet Karim for his support in coordinating with MH Samorita Medical College and Hospital to obtain the ultrasound images and metadata for this dataset. The author gratefully acknowledges the contributions of the radiologists, sonographers, and staff at MH Samorita Medical College and Hospital, for their assistance in conducting the ultrasound examinations and providing access to the imaging data. Their efforts in patient care and technical support were invaluable in making this dataset possible. Ethics and Data Access Permissions: This dataset was collected with formal approval from the Institutional Ethical Review Board (IERB) of MH Samorita Medical College and Hospital, Dhaka, Bangladesh. The ethical clearance was obtained before data collection and written institutional permission was granted before sharing. All images were anonymized and de-identified before inclusion in this dataset, ensuring patient privacy and compliance with ethical research standards. ## License: This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0). Users are required to cite this dataset when using it in publications, research, or derivative works. This dataset is openly available for research and academic purposes, supporting reproducibility and transparency in medical AI research.

UIdataGB Gallblader Diseases Dataset (Dataset)

The dataset is composed of ultrasound images of the GB organ from inside the gastrointestinal tract. The dataset includes 9 classes according to anatomical landmarks. Each class represents a GB disease. Published: 23 January 2024 | Version 1 | DOI: 10.17632/r6h24d2d3y.1 Turki, Amina; Mahdi Obaid, Ahmed; Bellaaj, Hatem; Ksantini, Mohamed; Altaee, Abdulla (2024), “Gallblader Diseases Dataset ”, Mendeley Data, V1, doi: 10.17632/r6h24d2d3y.1 "The UIdataGB dataset consists of 10692 images, annotated, and verified by medical doctorsand experienced radiologists. It includes 9 classes according to anatomical landmarks. Each classcontains nearly 1200 images. Therefore, the dataset is balanced in terms of diseases. In total,1782 patients were involved in the data collection; the number of female images was 6246,with an average age of 63.4, while the number of male images was 4446, with an average ageof 59.6.The number of images is sufficient to be used for different tasks, e.g., image retrieval, ML, DL,and transfer learning (TL), etc. The anatomical landmark of the GB determines the pathologicalfinding like cholecystitis, stone of the GB and polyps.The dataset consists of images with a resolution of 90 0×120 0 pixels and they are sorted intoseparate nine folders named according to the content. Tables 1 and 2 show the distribution ofdiseases in terms of images and patients’ numbers as well as the distribution of images accord-ing to gender."

EchoCP_dataset (Dataset)

A dataset of contrast transthoracic echocardiography, EchoCP, for patent foramen ovale diagnosis is published. We present EchoCP, the first dataset for cTTE based PFO diagnosis. EchoCP contains both VM and rest echocardiography videos captured from 30 patients. Data annotation including diagnosis annotation and segmentation annotation are performed by four experienced cardiovascular sonographers. As there are more than a thousand images in each patient s video, sparse labeling (only select representative frames) of the segmentation is adopted. EchoCP contains cTTE videos from 30 patients. For each patient, two videos corresponding to the rest and VM state of the patients are captured. Note that in the rest state, patients just relax and breathe normally. While in the VM, patients need to close their mouths, pinch their noses shut while expelling air out as if blowing up a balloon. The video is captured in the apical-4-chamber view and contains at least ten cardiac cycles. For the VM state, the action is performed three to five times during acquisition, and we selected the most representative one. If you used our dataset, please consider to cite our paper in MICCAI 2021, Tianchen Wang, Zhihe Li, Shanshan Bi, Meiping Huang, Jiawei Zhang, Jian Zhuang, Yiyu Shi, Hongwen Fei, Xiaowei Xu, "ImageCHD: A 3D Computed Tomography Image Dataset for Classification of Congenital Heart Disease," in Proc. of Medical Image Computing and Computer Assisted Interventions (MICCAI), Online, 2021. HIGHLIGHT 20231101: We have deployed the dataset on Kaggle! Please send emails to me xiao.wei.xu@foxmail.com if you have any questions about the dataset and the benchmark.

cardiacUDC_dataset (Dataset)

We collect CardiacUDA from our two hospitals: site G and site R. In order to guarantee all echocardiogram videos are standardscompliant, all cases of CardiacUDA are collected, annotated and approved by 5-6 experienced physicians. For ethical issues, we have required approval from medical institutions. Each patient underwent four views during scanning, which included parasternal left ventricle long axis (LVLA), pulmonary artery long axis (PALA), left ventricular short-axis (LVSA), and apical fourchamber heart (A4C), resulting in four videos per patient. The resolution of each video was either 800x600 or 1024x768, depending on the scanner used (Philips or HITACHI). A total of 516 and 476 videos were collected from Site G and Site R, respectively, from approximately 100 different patients. Each video consists of over 100 frames, covering at least one heartbeat cycle. We have provided pixel-level annotations for each view, including masks for the left ventricle (LV) and right ventricle (RV) in the LVLA view, masks for the pulmonary artery (PA) in the PALA view, masks for the LV and RV in the LVSA view, and masks for the LV, RV, left atrium (LA), and right atrium (RA) in the A4C view. The videos in both Site R and Site G were divided into a ratio of 8:1:1 for training, validation, and testing, respectively. To lower annotation costs in the training set, only five frames per video are provided with pixellevel annotation masks. To better measure the model performance, we provide pixel-level annotations for every frame in each video in the validation and testing sets. HIGHLIGHT 20231101: We have deployed the dataset on Kaggle! Please refer to the code () and our ICCV paper () for more detailes. Please send emails to me xiao.wei.xu@foxmail.com if you have any questions about the dataset and the benchmark.

ImperialAIEchocardiographyDataset_2020-12-05 (Dataset)

This is the latest versions of the datasets and code. They are constantly being added to. The code lives on github. Download 2020-12-05 release: Unity Imaging Echocardiography Model Development Dataset Images: Download Unity Imaging Echocardiography Model Development Dataset Labels: Download Unity Imaging Code: For reproducibility, specific snapshots of the datasets and code used for publication are below. Images - png-cache.zip 1) We curate a collection of DICOM files that will contribute to a dataset. 2) Each DICOM file is assigned to a dataset class - currently there are two 01 - development - training / tuning / internal validation images 02 - external validation images 3) Each DICOM file is given a 64 character hexadecimal code, e.g. 4d44413619e0161c5ab795bc1b899f7fb4bd0b2f5ab2efc881ecfc663d3bfb66 4) Each image within a DICOM (typically an individual frame for echo) gets given a number padded to 4 digits, starting from 0000 and going to 9999. 5) These images are extracted from the DICOM file, burnt-in meta-data masked, and saved as a png with their code as a filename - e.g. 01-4d44413619e0161c5ab795bc1b899f7fb4bd0b2f5ab2efc881ecfc663d3bfb66-0000.png 6) The individual images that make up a dataset for a paper are saved in a folder called png-cache, with sub directories for the dataset class (e.g. /01) and then the first two pairs of hexadecimal digits (e.g. /4d/44), i.e. /png-cache/01/4d/44/4d44413619e0161c5ab795bc1b899f7fb4bd0b2f5ab2efc881ecfc663d3bfb66-0000.png 7) This folder is then compressed to form png-cache.zip Not all files may have an associated label - e.g. all the frames of a video may be included, but only a few of them have expert labels Labels - labels.zip These are stored as JSON files. The development dataset (provided as labels-all.json) is divided up into: labels-train.json - training labels-tune.json - tuning labels-ival.json - internal validation For each image file (which acts as the key), there is a dictionary for every possible label. Each label for an image may have a type of: "off": the structure is definitely not in the image - i.e the outputs would be expected to be all zeros "blurred": the structure is might be in the image, but there is no label available (either it was too blurry, or no one has tried to label it) - i.e the output would need to be masked from the loss function "point": the structure is a single point, with the x and y coordinate from the x and y keys "curve": the structure is a curve, repreesnted as a cubic spline, with the x and y coordinates of the control points in the x and y keys For convenience each of the .json files have an equivalent .txt file with a list of the contained images.

Cardiac Assessment and Classification of Ultrasound (CACTUS) dataset (Dataset)

The Cardiac Assessment and Classification of Ultrasound (CACTUS) dataset is an open-graded dataset designed for the evaluation and classification of cardiac ultrasound images. The dataset was created as part of the ARQUS project, which aims to develop an autonomous robotic system capable of performing ultrasound scans and extracting quantitative measurements. This project is funded by the NSERC (Natural Sciences and Engineering Research Council of Canada). The dataset contains ultrasound images obtained from scans of the CAE Blue Phantom, a synthetic model used to simulate the human heart. These images represent a variety of heart views and exhibit different quality levels. A detailed grading schema was developed by two medical imaging experts to assess the quality of each image, which ensures that the dataset contains a diverse range of both high- and low-quality ultrasound scans. The CACTUS dataset is particularly valuable for applications in artificial intelligence, specifically in the domain of echocardiography. It has been used in the development of automated system for the classification of cardiac ultrasound images and the assessment of image quality, which can assist medical practitioners in automating these traditionally labor-intensive tasks. 1. Title of Dataset: CACTUS: An open dataset and framework for automated Cardiac Assessment and Classification of Ultrasound images using deep transfer learning. 2. Author Information A. Principal Investigator Contact Information Name: Hanae Elmekki Institution: Concordia University Email: hanae.elmekki@mail.concordia.ca B. Associate or Co-investigator Contact Information Name: Amanda Spilkin Institution: Concordia University Email: amanda.spilkin@mail.concordia.ca 3. Date of data collection (single date, range, approximate date): 2025-03-05 4. Geographic location of data collection: Concordia University, Montreal, Quebec, Canada 5. Information about funding sources that supported the collection of the data: Natural Sciences and Engineering Research Council of Canada (NSERC), Discovery Horizons Program and Individual Discovery Grant —————————————- SHARING/ACCESS INFORMATION —————————————- 1. Licenses/restrictions placed on the data: These data are available under a CC BY 4.0 license <> 2. Links to publications that cite or use the data:

EchoUNAL a benchmarking study for automatic classification of six echocardiographic views (Dataset)

# EchoUNAL Dataset & Classifier This repository contains the dataset and resources for the paper, "The EchoCardiography open data base EchoUNAL: a benchmarking study for automatic classification of six echocardiographic views". ## Description The study presents a comparative evaluation of four CNN architectures for classifying six standard echocardiographic views: * **A4C**: Apical Four Chamber * **A5C**: Apical Five Chamber * **PLAX**: Parasternal Long Axis * **PSAX**: Parasternal Short Axis * **S4C**: Subcostal Four Chambers * **IVC**: Inferior Vena Cava ## Dataset Details The database was created specifically for this study and is shared publicly to encourage further research. * **Data Source**: The dataset was obtained from **89 healthy volunteers**, with informed consent from all participants. * **Acquisition Hardware**: A **Butterfly iQ+** ultrasound device was used for image acquisition. * **Acquisition Protocol**: Images were captured by two cardiologists specializing in echocardiography. Some videos were excluded due to poor acoustic window quality. * **Total Contents**: The final dataset comprises **346 videos**, distributed as follows: * **A4C**: 61 videos * **A5C**: 49 videos * **PLAX**: 66 videos * **PSAX**: 61 videos * **S4C**: 54 videos * **IVC**: 55 videos ## Data Collection & Preprocessing * **Original Format**: Videos were stored in .AVI format with a 500x500 pixel resolution. The original frame rate varied between 26 and 30 fps. * **Applied Preprocessing**: 1. The ultrasound interface was cropped to center the cardiac image area. 2. The frame rate was standardized to **30 fps** using interpolation. 3. A manual, frame-level relabeling was performed to ensure label accuracy. A **"no view"** label was used for frames without a clear view, though this class was later excluded from the training setup due to its high heterogeneity. 4. All video frames were resized to **224x224** pixels, a standard input size for the pretrained CNN architectures. ## Acknowledgments This work was funded by the project "Estimation of cardiac work as an index of cardiovascular function in echocardiographic videos" with code 60946 from the call for research projects SUE Distrito Capital of 2023. J. D. S. Avila et al., "The EchoCardiography open data base EchoUNAL: a benchmarking study for automatic classification of six echocardiographic views," 2025 21st International Symposium on Biomedical Image Processing and Analysis (SIPAIM), Pasto, Colombia, 2025, pp. 1-4, doi: 10.1109/SIPAIM67325.2025.11283340. keywords: Training;Visualization;Ultrasonic imaging;Systematics;Echocardiography;Computational modeling;Benchmark testing;Convolutional neural networks;Videos;Residual neural networks

IUGC Ultrasound Dataset (MICCAI 2025) (Dataset)

Lung Ultrasound Dataset (LUS-Dataset-Katumba) (Dataset)

This dataset contains a curated benchmark collection of 1,062 labelled lung ultrasound (LUS) images collected from patients at Mulago National Referral Hospital and Kiruddu Referral Hospital in Kampala, Uganda. The images were acquired and annotated by senior radiologists to support the development and evaluation of artificial intelligence (AI) models for pulmonary disease diagnosis. Each image is categorized into one of three classes: Probably COVID-19 (COVID-19), Diseased Lung but Probably Not COVID-19 (Other Lung Disease), and Healthy Lung. The dataset addresses key challenges in LUS interpretation, including inter-operator variability, low signal-to-noise ratios, and reliance on expert sonographers. It is particularly suitable for training and testing convolutional neural network (CNN)-based models for medical image classification tasks in low-resource settings. The images are provided in standard formats such as PNG or JPEG, with corresponding labels stored in structured files like CSV or JSON to facilitate ease of use in machine learning workflows. In this second version of the dataset, we have extended the resource by including a folder containing the original unprocessed raw data, as well as the scripts used to process, clean, and sort the data into the final labelled set. These additions promote transparency and reproducibility, allowing researchers to understand the full data pipeline and adapt it for their own applications. This resource is intended to advance research in deep learning for lung ultrasound analysis and to contribute toward building more accessible and reliable diagnostic tools in global health. Katumba, Andrew; Murindanyi, Sudi; Okila, Nixson; Nakatumba-Nabende, Joyce; Mwikirize, Cosmas; Serugunda, Jonathan; Bugeza, Samuel; Oriekot, Anthony; Bosa, Juliet; Nabawanuka, Eva (2025), “A Dataset of Lung Ultrasound Images for Automated AI-based Lung Disease Classification”, Mendeley Data, V2, doi: 10.17632/hb3p34ytvx.2

OpenPOCUS - Lung Ultrasound Image Database (Dataset)

Background Lung ultrasound (LUS) offers advantages over traditional imaging for diagnosing pulmonary conditions, with superior accuracy compared to chest X-ray and similar performance to CT at lower cost. Despite these benefits, widespread adoption is limited by operator dependency, moderate interrater reliability, and training requirements. Deep learning (DL) could potentially address these challenges, but development of effective algorithms is hindered by the scarcity of comprehensive image repositories with proper metadata.

SegThy Open-Access Dataset for Thyroid and Neck Segmentation (Dataset)

## Motivation Ultrasound (US) imaging plays a central role in the diagnosis of thyroid diseases as well as different pathologies of the neck region. Additionally, US is used for treatment planning, in the case of radioiodine therapy, and also as mean of following up on the success of different therapeutic efforts. Yet, in the vast majority of hospitals, 2D freehand US is applied. This type of examination has shown to have a high intra-observer and inter-observer variability [1], and low accuracy in terms of volume prediction in a thyroid volumetry setting using MRI as ground truth. To tackle these problems, we propose the use of 3D US in combination with machine learning (ML) algorithms to automatically segment the most relevant organs in the region, as well as anomalies, such as nodules and tumors. 3D US can be acquired in several ways, e.g. using 3D US probes, robotic US, so-called wobbler US probes, or using freehand tracked US. The last option has gained traction over the last years as it enables to extend almost any existing commercial 2D US at a low cost. On the side of ML, its availability and increasing computing power are flooding the medical world. Yet, ML can only provide trustworthy results if sufficient (annotated) data is available to “learn” from it. This motivated us to create and publish this dataset, and thus make a relevant contribution to the community. ## Citation [1] Tracked 3D ultrasound and deep neural network-based thyroid segmentation reduce interobserver variability in thyroid volumetry; M. Krönke, C. Eilers, D. Dimova, M. Köhler, G. Buschner, L. Schweiger, L. Konstantinidou, M. Makowski, J. Nagarajah, N. Navab, W. Weber, and T. Wendler; PLOS ONE - July 29, 2022 - ## Dataset Description Sub-dataset 1 This sub-dataset consists of 28 (healthy) volunteers who were scanned using freehand tracked ultrasonography (both sides of their neck, focus on thyroid). A Siemens Acuson NX-3 US machine, combined with a 12MHz VF12-4 probe, was adapted to be tracked using electromagnetic tracking using the Piur tUS system. Additionally, all patients were imaged with MRI (Siemens Biograph mMR) using a T1-weighted VIBE (volumetric interpolated breath-hold) sequence centered in the thyroid area of the neck. The magnetic field was set to 3T. In practical terms, each volunteer in this subdataset has: 1x MRI (T1-weighted VIBE sequence) of the neck area, with voxel size 0.625x0.625x1 mm3 and field of view 320x320x80 mm3. 9x two-sided 3D US covering the thyroid region with voxel size of 0.12x0.12x0.12 mm3 and variable field of view. The nine 3D US were acquired by three physicians (three scans each). Label maps for the thyroid in all MR and US images. Sub-dataset 2 The second subdataset consists of 186 patients undergoing routinary thyroid US either as initial diagnostics means or as follow-up. The same US device and freehand tracked US extension device were used. Each patient in this sub-dataset has 6x two-sided 3D US covering the thyroid. The 6 3D US were acquired by 2 physicians (three scans each). In both dataset, the age and the sex of the patients is included in the metadata. The dataset is currently being extended to include the labels of the trachea, the common carotid arteries, the internal jugular veins, and (if present) thyroid nodules in sub-dataset 1 (in both US and MRI), and of the same organs in sub-dataset 2 in US. See figures as examples for the multilabel annotations. ## License This dataset is licensed under a CC BY license. This means that users of it can distribute, shuffle, adapt, and build upon the material with the sole condition that the creators are acknowledged. The license permits commercial use as long as the authors are cited.

CAMUS Cardiac Acquisitions for Multi-structure Ultrasound Segmentation (Dataset)

The goal of this project is to provide all the materials to the community to resolve the problem of echocardiographic image segmentation and volume estimation from 2D ultrasound sequences (both two and four-chamber views). To this aim, the following solutions were set-up introduction of the largest publicly-available and fully-annotated dataset for 2D echocardiographic assessment (to our knowledge). The CAMUS dataset, containing 2D apical four-chamber and two-chamber view sequences acquired from 500 patients, is made available for download # Dataset properties The overall CAMUS dataset consists of clinical exams from 500 patients, acquired at the University Hospital of St Etienne (France) and included in this study within the regulation set by the local ethical committee of the hospital after full anonymization. The acquisitions were optimized to perform left ventricle ejection fraction measurements. In order to enforce clinical realism, neither prerequisite nor data selection have been performed. Consequently, - some cases were difficult to trace; - the dataset involves a wide variability of acquisition settings; - for some patients, parts of the wall were not visible in the images; for some cases, the probe orientation recommendation to acquire a rigorous four-chambers view was simply impossible to follow and a five-chambers view was acquired instead. This produced a highly heterogeneous dataset, both in terms of image quality and pathological cases, which is typical of daily clinical practice data. The dataset has been made available to the community HERE. The dataset comprises : i) a training set of 450 patients along with the corresponding manual references based on the analysis of one clinical expert; ii) a testing set composed of 50 new patients. The raw input images are provided through the raw/mhd file format. # Study population Half of the dataset population has a left ventricle ejection fraction lower than 45%, thus being considered at pathological risk (beyond the uncertainty of the measurement). Also, 19% of the images have a poor quality (based on the opinion of one expert), indicating that for this subgroup the localization of the left ventricle endocarium and left ventricle epicardium as well as the estimation of clinical indices are not considered clinically accurate and workable. In classical analysis, poor quality images are usually removed from the dataset because of their clinical uselessness. Therefore, those data were not involved in this project during the computation of the different metrics but were used to study their influence as part of the training and validation sets for deep learning techniques. # Involved systems

HMC-QU echocardiography ultrasound recordings (Dataset)

The HMC-QU benchmark dataset is created by the collaboration between Hamad Medical Corporation (HMC), Tampere University, and Qatar University. The usage of data has been approved by the local ethics board of HMC Hospital in February 2019. The dataset includes a collection of apical 4-chamber (A4C) and apical 2-chamber (A2C) view 2D echocardiography recordings obtained during the years 2018 and 2019. The echocardiography recordings are acquired via devices from different vendors that are Phillips and GE Vivid (GE-Health-USA) ultrasound machines. The temporal resolution (frame rate per second) of the echocardiography recordings is 25 fps. The spatial resolution varies from 422x636 to 768x1024 pixels. The dataset can be utilized for both myocardial infarction (heart attack) detection and left ventricle wall segmentation purposes. # Detection of Myocardial Infarction HMC-QU is the first dataset that is shared with the research community serving myocardial infarction (MI) detection on the left ventricle wall of the heart. The recordings are from over 10,000 echos performed in a year including more than 800 cases admitted with acute ST-elevation MI. The patients with MI were treated with coronary angiogram/angioplasty after the diagnosis of acute MI with electrocardiography and cardiac enzymes evidence. The patients had echocardiography recordings obtained within 24 hours of admission or in some cases before they underwent coronary angioplasty. The subjects not diagnosed with MI underwent a required health check and investigation for other reasons in the hospital. The ground-truth labels are provided for each myocardial segment illustrated in Figure 1 as non-MI and MI, where the MI term indicates any sign of regional wall motion abnormality, whereas the subjects without regional wall motion abnormality are assigned to non-MI. The one cardiac cycle frames are predefined for each recording. End-diastole and end-systole frames are defined according to the electrocardiography (ECG) recordings of the patients. For the patients without ECG recordings, the cardiac cycle is defined according to the frames, where the left ventricle area is the largest and smallest. 1.1. Apical 4-chamber HMC-QU dataset consists of 162 A4C view 2D echocardiography recordings. The A4C view recordings belong to 93 MI patients (all first-time and acute MI) and 69 non-MI subjects. 1.2. Apical 2-chamber The dataset consists of 130 A2C view 2D echocardiography recordings that belong to 68 MI patients and 62 non-MI subjects. # Segmentation of the Left Ventricle Wall A subset of 109 A4C view echocardiography recordings has their corresponding ground-truth segmentation masks for the whole left ventricle wall at each frame for one cardiac cycle. This subset includes 72 MI patients and 37 non-MI subjects. The size of the ground-truth segmentation masks is 224x224 in order to have suitable input dimensions for many state-of-the-art deep network topologies. If you use the HMC-QU dataset in your research, please consider citing the publications below: [P1] A. Degerli, S. Kiranyaz, T. Hamid, R. Mazhar, and M. Gabbouj, “Early Myocardial Infarction Detection over Multi-view Echocardiography,” arXiv preprint arXiv:2111.05790v2, 2021, . [P2] A. Degerli, M. Zabihi, S. Kiranyaz, T. Hamid, R. Mazhar, R. Hamila, and M. Gabbouj, "Early Detection of Myocardial Infarction in Low-Quality Echocardiography," in IEEE Access, vol. 9, pp. 34442-34453, 2021, . [P3] S. Kiranyaz, A. Degerli, T. Hamid, R. Mazhar, R. E. F. Ahmed, R. Abouhasera, M. Zabihi, J. Malik, R. Hamila, and M. Gabbouj, "Left Ventricular Wall Motion Estimation by Active Polynomials for Acute Myocardial Infarction Detection," in IEEE Access, vol. 8, pp. 210301-210317, 2020, .

Breast Ultrasound Images Dataset (Dataset BUSI) (Dataset)

The data collected at baseline include breast ultrasound images among women in ages between 25 and 75 years old. This data was collected in 2018. The number of patients is 600 female patients. The dataset consists of 780 images with an average image size of 500*500 pixels. The images are in PNG format. The ground truth images are presented with original images. The images are categorized into three classes, which are normal, benign, and malignant. If you use this dataset, please cite: Al-Dhabyani W, Gomaa M, Khaled H, Fahmy A. Dataset of breast ultrasound images. Data in Brief. 2020 Feb;28:104863. DOI: 10.1016/j.dib.2019.104863. | Subject area | Medicine and Dentistry | |——————————————|——————————————————————————————————————————————————————————————————————————————————————————| | More specific subject area | Radiology and Imaging | | Type of data | Images and mask images | | How data was acquired | LOGIQ E9 ultrasound and LOGIQ E9 Agile ultrasound system | | Data format | PNG | | Experimental factors | All images are classified as normal, benign and malignant | | Experimental features | When medical images are used for training deep learning models, they provide fast and accurate results in classification, detection, and segmentation of breast cancer. | | Data source location | Baheya Hospital for Early Detection & Treatment of Women s Cancer, Cairo, Egypt. |

2D ultrasound sequences of the liver (mp4) (Dataset)

7 2D ultrasound sequences of the liver of healthy volunteers were acquired during free breathing over a period of 5-10 min. This is a converted version of the usliverseq dataset into mp4 files encoded with H.264 to reduce size and make the files easier to read. The conversion commands used are: