Deltadahl/OCT-Longitudinal

Dataset Card for OCT-Longitudinal

Overview

This dataset comprises 1.1 million synthetic OCT images paired with corresponding synthetic longitudinal data, specifically designed for the development and testing of machine learning models in the medical imaging domain.

Dataset Description

The OCT-Longitudinal dataset facilitates the exploration of the informative value of images in predicting longitudinal patient outcomes. A structured latent space created by a Variational Autoencoder (VAE) trained on a public dataset of 109,309 OCT images guides the Stable Diffusion (SD) model to generate new images. These images are linked to synthetic longitudinal data generated through a Nonlinear Mixed-Effects (NLME) model, where the noise levels in data vary to represent explanatory values ranging from 100% to 2%.

• —Curated by: Simon Deltadahl

• —License: cc-by-4.0
• —Repository: Available at GitHub and Hugging Face <!-- - Paper: TODO -->

Repository Structure

OCT-Longitudinal/
│
├── 128_dim_latent_space/             # The latent space vectors that were used to create the images and the longitudinal data

│
├── longitudinal_data/                # The longitudinal data
│
├── generated_images/                 # Generated images by the SD model
│   ├── 100k_test/                    # The 100,000 test set
│   └── 1M_train/                     # The 1,000,000 train and validation set
│
├── saved_models/                     # The trained models used in the paper
│   ├── SD_OCT/                       # The Stable Diffusion model
│   └── VAE.jld2/                     # The Variational Auto Encoder
│
├── test_images/                      # The images that were used as the test set
│
├── croissant.json/                   # Metadata
│
└── .gitignore                        # Specifies intentionally untracked files to ignore

Intended Use

This dataset is intended for academic and research purposes, particularly for:

• —Benchmarking predictive algorithms that estimate patient outcomes from medical images.
• —Analyzing the impact of varying informative values of images on algorithm performance.
• —Exploring the synthetic correlation between images and longitudinal data in controlled settings.

Out-of-Scope Use

The dataset is not designed for:

• —Direct clinical application without real-world data validation.
• —Legal, diagnostic, or treatment purposes.
• —Training models without acknowledging the synthetic nature of the data.

Dataset Structure

Curation Rationale

The dataset was created to provide a privacy-compliant, comprehensive resource to precisely adjust the informative value of medical images relative to outcomes. This assists in overcoming the challenges of benchmarking predictive models without exact knowledge of data explainability due to uncertainties inherent to medical data.

Source Data

Data Collection and Processing

The data was synthesized using a combination of VAE, SD, and NLME models, with noise levels introduced to control the correlation between the generated images and longitudinal data.

Data Producers

The synthetic data was produced by Simon Deltadahl, Andreu Vall, Vijay Ivaturi, and Niklas Korsbo.

Bias, Risks, and Limitations

Due to its synthetic nature, this dataset might not capture the full complexity or variability of real-world medical data. Users should be cautious about the direct applicability of their findings without further validation.

Recommendations

Researchers are advised to: - Validate findings with real-world data.

• —Clearly state the synthetic nature of the data when publishing results.
• —Primarily use this dataset for preliminary testing and model development.

Citation

BibTeX:

@misc{simon_deltadahl_2024,
  author       = {Deltadahl, Simon and Vall, Andreu and Ivaturi, Vijay and Korsbo, Niklas},
  title        = {OCT-Longitudinal},
  year         = 2024,
  url          = {https://huggingface.co/datasets/Deltadahl/OCT-Longitudinal},
  doi          = {10.57967/hf/2089},
  publisher    = {Hugging Face}
}

Authors and Contact

• —Dataset Card Author: Simon Deltadahl
• —Contact: For more information, please contact Simon Deltadahl at scfc3@cam.ac.uk

Datasheet

Motivation

1. 1.For what purpose was the dataset created?

The dataset was created for academic and research purposes, particularly to enable the benchmarking of predictive algorithms that estimate patient outcomes from medical images. It aims to analyze the impact of varying informative values of images on algorithm performance and explore synthetic correlations between images and longitudinal data in controlled settings. The dataset was designed to fill the gap in openly available, large-scale, and synthetic patient data that can be used for developing and testing machine learning models in healthcare research without the privacy concerns associated with real patient data.

1. 1.Who created the dataset?

The dataset was created by Simon Deltadahl from the Department of Applied Mathematics and Theoretical Physics at the University of Cambridge, and Andreu Vall, Vijay Ivaturi, and Niklas Korsbo from Pumas-AI Inc.

Composition

1. 1.What do the instances that comprise the dataset represent (e.g., documents, photos, people, countries)? Are there multiple types of instances (e.g., movies, users, and ratings; people and interactions between them; nodes and edges)? Please provide a description.

The instances in the dataset represent synthetic medical data, including optical coherence tomography (OCT) scan images and corresponding longitudinal patient observations. There are multiple types of instances: images and associated longitudinal data points.

1. 1.How many instances are there in total?

The dataset comprises 1.1 million OCT scan images paired with five sets of longitudinal observations, totaling 5.5 million longitudinal data pairs.

1. 1.Does the dataset contain all possible instances or is it a sample (not necessarily random) of instances from a larger set? If the dataset is a sample, then what is the larger set? Is the sample representative of the larger set (e.g., geographic coverage)? If so, please describe how this representativeness was validated/verified. If it is not representative of the larger set, please describe why not (e.g., to cover a more diverse range of instances, because instances were withheld or unavailable).

The dataset is a synthetic sample created to explore varying levels of correlation between images and longitudinal data. It is not representative of a real-world patient population but is designed to cover a wide range of controlled associations for benchmarking purposes.

1. 1.What data does each instance consist of? "Raw" data (e.g., unprocessed text or images) or features? In either case, please provide a description.

Each instance consists of an OCT scan image in JPEG format and longitudinal observation data stored in JSON format. The images represent the structural information of the eye, while the longitudinal data include synthetic patient outcomes over time.

1. 1.Is there a label or target associated with each instance? If so, please provide a description.

Yes, each OCT scan image is associated with longitudinal observation targets that represent synthetic patient outcomes over time.

1. 1.Is any information missing from individual instances? If so, please provide a description, explaining why this information is missing (e.g., because it was unavailable). This does not include intentionally removed information, but might include, e.g., redacted text.

No information is missing from the instances. Each instance is complete with both the image and its corresponding longitudinal data.

1. 1.Are relationships between individual instances made explicit (e.g., users’ movie ratings, social network links)? If so, please describe how these relationships are made explicit.

Yes, relationships between individual instances are made explicit by linking each OCT scan image with its corresponding longitudinal observations. This linkage is maintained through consistent indexing.

1. 1.Are there recommended data splits (e.g., training, development/validation, testing)? If so, please provide a description of these splits, explaining the rationale behind them.

We provide a train/test split of 1,000,000 training instances and 100,000 test instances. However, users are free to split the training set into a train/validation split according to their experimental needs.

1. 1.Are there any errors, sources of noise, or redundancies in the dataset? If so, please provide a description.

The dataset intentionally introduces controlled noise in the longitudinal observations to create different levels of correlation with the images. This is to test the robustness of predictive models under varying levels of signal strength.

1. 1.Is the dataset self-contained, or does it link to or otherwise rely on external resources (e.g., websites, tweets, other datasets)? If it links to or relies on external resources, a) are there guarantees that they will exist, and remain constant, over time; b) are there official archival versions of the complete dataset (i.e., including the external resources as they existed at the time the dataset was created); c) are there any restrictions (e.g., licenses, fees) associated with any of the external resources that might apply to a dataset consumer? Please provide descriptions of all external resources and any restrictions associated with them, as well as links or other access points, as appropriate.

The dataset is self-contained and does not rely on external resources.

1. 1.Does the dataset contain data that might be considered confidential (e.g., data that is protected by legal privilege or by doctor–patient confidentiality, data that includes the content of individuals’ non-public communications)? If so, please provide a description.

No, the dataset contains synthetic data generated for research purposes and does not include any confidential or real patient information.

1. 1.Does the dataset contain data that, if viewed directly, might be offensive, insulting, threatening, or might otherwise cause anxiety? If so, please describe why.

No, the dataset does not contain any content that might be offensive, insulting, threatening, or cause anxiety.

1. 1.Does the dataset identify any subpopulations (e.g., by age, gender)? If so, please describe how these subpopulations are identified and provide a description of their respective distributions within the dataset.

No, the dataset does not identify any subpopulations. All data are synthetic and do not include demographic identifiers such as age or gender.

1. 1.Is it possible to identify individuals (i.e., one or more natural persons), either directly or indirectly (i.e., in combination with other data) from the dataset? If so, please describe how.

No, it is not possible to identify individuals from the dataset. The data are entirely synthetic and anonymized.

1. 1.Does the dataset contain data that might be considered sensitive in any way (e.g., data that reveals race or ethnic origins, sexual orientations, religious beliefs, political opinions or union memberships, or locations; financial or health data; biometric or genetic data; forms of government identification, such as social security numbers; criminal history)? If so, please provide a description.

No, the dataset does not contain any sensitive data. All instances are synthetic and do not include any real or sensitive personal information.

1. 1.Any other comments?

The dataset is intended solely for academic and research purposes to explore the relationship between medical images and longitudinal data. It is crucial for users to acknowledge the synthetic nature of the dataset and its limitations for real-world applications.

Collection Process

1. 1.How was the data associated with each instance acquired? Was the data directly observable (e.g., raw text, movie ratings), reported by subjects (e.g., survey responses), or indirectly inferred/derived from other data (e.g., part-of-speech tags, model-based guesses for age or language)? If the data was reported by subjects or indirectly inferred/derived from other data, was the data validated/verified? If so, please describe how.

The data was acquired by generating synthetic images using a Variational Autoencoder (VAE) combined with a Stable Diffusion (SD) model, and synthetic longitudinal observations were created using a Nonlinear Mixed Effects (NLME) model. The models were trained on a public dataset of 109,309 2D Optical Coherence Tomography (OCT) scan slices, and the longitudinal data was simulated from a low-dimensional space of random effects.

1. 1.What mechanisms or procedures were used to collect the data (e.g., hardware apparatuses or sensors, manual human curation, software programs, software APIs)? How were these mechanisms or procedures validated?

The synthetic data generation involved the use of software programs (Python and Julia) for training the VAE and SD models to generate images and the NLME model to simulate longitudinal data. These mechanisms were validated through the evaluation of the generated data's predictive capabilities in empirical tests.

1. 1.If the dataset is a sample from a larger set, what was the sampling strategy (e.g., deterministic, probabilistic with specific sampling probabilities)?

The dataset is not a sample from a larger set but rather entirely synthesized. The VAE and SD models were trained on a large, publicly available dataset of OCT images.

1. 1.Did you collect the data from the individuals in question directly, or obtain it via third parties or other sources (e.g., websites)?

The data used to train the models were obtained from publicly available sources, specifically from the OCT dataset licensed under CC BY 4.0.

Preprocessing/Cleaning/Labeling

1. 1.Was any preprocessing/cleaning/labeling of the data done (e.g., discretization or bucketing, tokenization, part-of-speech tagging, SIFT feature extraction, removal of instances, processing of missing values)? If so, please provide a description.

Yes, preprocessing was conducted on the data:

• —Image Data: Optical coherence tomography (OCT) images were preprocessed by resizing them to ensure consistent input dimensions for model training. The white background of the images was replaced with the mean value of the non-white background.
• —Labeling: Labels for the longitudinal data were generated synthetically by the NLME model, ensuring controlled levels of association with the corresponding OCT images.

1. 1.Is the software that was used to preprocess/clean/label the data available? If so, please provide a link or other access point.

Yes, the software used for preprocessing and cleaning is available. The code can be accessed from the following repository: https://github.com/Deltadahl/Image-Longitudinal/tree/main/src/data_manipulation.

Uses

1. 1.What (other) tasks could the dataset be used for?

The dataset could be used for:

• —Developing and benchmarking new predictive models for medical outcomes.
• —Testing the impact of synthetic data on the training and performance of machine learning algorithms.
• —Studying the relationship between medical imaging data and longitudinal patient outcomes.
• —Exploring novel methods for generating synthetic medical data.

1. 1.Is there anything about the composition of the dataset or the way it was collected and preprocessed/cleaned/labeled that might impact future uses?

Yes, the synthetic nature of the data should be carefully considered. Using this dataset without acknowledging its synthetic nature could lead to misinterpretation of results, especially in clinical settings. Additionally, the dataset was generated with controlled associations between images and longitudinal data, which may not represent real-world variability.

1. 1.Are there tasks for which the dataset should not be used?

Yes, the dataset should not be used for:

• —Direct clinical applications, diagnostics, or treatment decisions.
• —Legal purposes or any scenarios that require real patient data.
• —Training models without accounting for the synthetic nature of the data.

1. 1.Any other comments?

It is crucial to ensure that users of the dataset are aware of its intended research and development purposes and the potential limitations due to its synthetic origin.

Distribution

1. 1.Will the dataset be distributed to third parties outside of the entity (e.g., company, institution, organization) on behalf of which the dataset was created?

Yes, the dataset is publicly available on the internet.

1. 1.How will the dataset be distributed (e.g., tarball on website, API, GitHub)? Does the dataset have a digital object identifier (DOI)?

The dataset is distributed on Hugging Face at https://doi.org/10.57967/hf/2089. The dataset does have a DOI.

1. 1.When will the dataset be distributed?

The dataset was first released in May 2024.

1. 1.Will the dataset be distributed under a copyright or other intellectual property (IP) license, and/or under applicable terms of use (ToU)? If so, please describe this license and/or ToU, and provide a link or other access point to, or otherwise reproduce, any relevant licensing terms or ToU, as well as any fees associated with these restrictions.

The dataset is licensed under CC BY 4.0, which allows others to share and adapt the material as long as appropriate credit is given. For more details, see the license at https://creativecommons.org/licenses/by/4.0/.

1. 1.Do any export controls or other regulatory restrictions apply to the dataset or to individual instances? If so, please describe these restrictions, and provide a link or other access point to, or otherwise reproduce, any supporting documentation.

No, there are no known export controls or other regulatory restrictions applicable to the dataset.

Maintenance

1. 1.Who will be supporting/hosting/maintaining the dataset?

The dataset will be supported and maintained by the authors. The dataset is hosted on Hugging Face.

1. 1.How can the owner/curator/manager of the dataset be contacted (e.g., email address)?

The owners and curators of the dataset can be contacted via the following email addresses:

• —Simon Deltadahl: scfc3@cam.ac.uk
• —Andreu Vall: andreu@pumas.ai
• —Vijay Ivaturi: vijay@pumas.ai
• —Niklas Korsbo: niklas@pumas.ai

You can also open a new discussion on our Hugging Face page.

1. 1.Is there an erratum? If so, please provide a link or other access point.

Any errata related to the dataset will be documented and updated on the dataset's page on Hugging Face at https://doi.org/10.57967/hf/2089.

1. 1.Any other comments?

The authors are committed to ensuring the long-term accessibility and maintenance of the dataset. They will monitor and address any issues related to data availability and user access promptly.