Vision Intelligence Network

Explainable machine learning for eye disease

Machine learning systems allow clinicians and researchers to resolve complex underpinnings of visual disorders, using high-dimensional inputs from RNA microarrays to OCT images. With the development of numerous trained models and curated datasets, eye-disease specialists require a systematic way to examine and implement these tools. The Vision Intelligence Network aims to provide an accessible database of machine learning tools for ophthalmology, improving accessibility, explainability, and generalizability in the field.

Methodology

The Vision Intelligence Network (VIN) applies natural language processing to survey scientific literature for the most recent published reports of machine learning methods in ophthalmology. The reports are processed to extract relevant hard-coded metadata and keywords. Additionally, each entry undergoes analysis with an LLM for unstructured summary information.

For each entry, the following attributes are cataloged:

• Condition: The primary pathology that the model investigates.
• Model name: Provided by the publication, otherwise defaults to author name.
  
• Description: Brief description of model, generally the title of the reporting publication.
• Task: Brief description of the task to which the model applies.*
• Task type: The primary learning operation (segmentation, forecasting, or detection/diagnosis), if available.
• Model type: Primary type of algorithm implemented, if available.
• Input data type: Description of the type of data used for training.*
• Training set size: Number of examples in the training set.*
• Inferred sensitivity: Reported sensitivity of the model in the primary evaluation task.*
• Inferred specificity: Reported specificity of the model in the primary evaluation task.*
• Original report: Link to the associated article.
• Code link: Link to the associated code, if available.
• Code availability: Whether corresponding code is publicly available.
• Publication date: For corresponding paper, if available.

* These data elements are identified from unstructured text by querying with a general-purpose LLM (Google Gemini Pro). They may be inaccurate, and users are recommended to consult the corresponding manuscript for details of model evaluation.

Contact

team@visintelnet.org

Disclaimer

This database, including its curation and organization, is experimental. It is intended entirely as a research tool to facilitate the implementation of academic methods. Any information displayed herein may be inaccurate or outdated. For information regarding any method described in this database, please refer to the respective citation.