Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

InSight: AI Mobile Screening Tool for Multiple Eye Disease Detection using Multimodal Fusion

Created by
  • Haebom

Author

Ananya Raghu, Anisha Raghu, Alice S. Tang, Yannis M. Paulus, Tyson N. Kim, Tomiko T. Oskotsky

Outline

In this paper, we present InSight, an AI-based mobile app for early diagnosis of five major ophthalmic diseases (age-related macular degeneration, glaucoma, diabetic retinopathy, diabetic macular edema, and pathological myopia) with limited accessibility in low- and middle-income countries and resource-poor settings. InSight diagnoses diseases by combining patient metadata with fundus images, and consists of a three-stage pipeline: real-time image quality assessment, a disease diagnosis model, and a diabetic retinopathy severity assessment model. The disease diagnosis model integrates three key innovations: a multi-modal fusion technique (MetaFusion) that combines metadata and images, a pre-training method utilizing supervised and self-supervised learning loss functions, and a multi-task model that simultaneously predicts the five diseases. The model is trained and evaluated using BRSET (lab-captured images) and mBRSET (smartphone-captured images) datasets, and shows high diagnostic accuracy under various image conditions captured in both smartphones and lab settings.

Takeaways, Limitations

•
Takeaways:
◦
Presenting the possibility of providing an effective AI-based solution for early diagnosis of eye diseases in areas with low access to healthcare, such as low- and middle-income countries.
◦
High diagnostic accuracy and efficient model building through multi-modal fusion technique (MetaFusion), pre-learning method, and multi-task model.
◦
It maintains high accuracy even in videos taken with a smartphone, increasing the possibility of actual field application.
◦
Achieve 5x improved computational efficiency compared to using 5 individual models.
•
Limitations:
◦
Lack of information on the specific size and composition of the BRSET and mBRSET datasets. Further validation of the datasets' bias and generalizability is needed.
◦
Further research is needed on actual clinical usefulness and effectiveness through clinical trials.
◦
Further analysis is needed on the model's explainability and reliability.
◦
Generalization performance evaluation across different races and age groups is needed.
👍
Made with Slashpage