Independent Highly-Accurate, And Interactive Ai-Based Search System For Publicly Available Anterior Segment Imaging Datasets

Building on the manually performed review of world-wide publically available anterior segment datasets published by authors in JCRS 2024, the aim of the study was to build a novel AI-driven platform capable of the same. 
Advanced natural language processing (NLP) and machine learning techniques were deployed to automate dataset identification, categorisation, and metadata harmonisation. Additional purpose was to develop an interactive AI agent to serve as a more natural interface than a webpage, streamlining user interactions and refining dataset exploration. By minimising the need for manual searches, we sought to enhance reproducibility, reduce resource demands, and improve the overall accuracy of curated datasets for ophthalmic research.

This work was conducted as an academically sponsored initiative under the auspices of the ESCRS Digital Health Special Interest Group. Data sources included both open-access and subscription-based platforms, analysed in line with institutional and GDPR guidelines. The AI-driven pipeline was complemented by an interactive AI agent, enabling researchers to query datasets conversationally, improving accessibility and usability.

We developed an AI-powered pipeline that performed structured searches across multiple academic repositories (PubMed, Google Dataset Search, et al). NLP algorithms refined search terms to capture diverse dataset metadata, while machine learning models automatically flagged dataset accessibility (open, restricted, or inaccessible). To enhance usability, an interactive AI agent was integrated, allowing researchers to interact conversationally with the system for dataset exploration and retrieval. A secondary manual review resolved ambiguities, ensuring consistent metadata formats for integration into a central repository. Performance was benchmarked against the manually curated review to assess improvements in discovery speed and data quality

The AI-based pipeline identified 150 unique anterior segment imaging datasets, including seven that had been previously overlooked. Concurrently, three earlier datasets were deemed inaccessible, underscoring the value of continuous monitoring. Compared to manual methods, the approach reduced discovery time by approximately 80%. The addition of an interactive AI agent improved user engagement, making dataset retrieval more intuitive and efficient. Standardised metadata facilitated cross-study comparisons and improved dataset diversity, capturing broader geographic sources. These results highlight the potential for AI to address the limitations of manual data curation in ophthalmology.

Our findings indicate that the integration of AI into dataset reviews significantly enhances speed, accuracy, and data comprehensiveness. The inclusion of an interactive AI agent provided a more natural and accessible interface, reducing the complexity of dataset discovery for researchers. By automating discovery and updating processes, AI minimises reliance on time-intensive manual curation, thereby improving reproducibility and accelerating research progress. This approach offers a scalable model that could be adopted across various medical disciplines, strengthening the global ophthalmic research ecosystem and aligning with FAIR data principles.