TLDR: InSight is an AI-powered mobile app that uses both eye images and patient health data to accurately detect five common eye diseases simultaneously. It includes a quality check for images, a smart diagnosis model that learns from various data, and a system to grade disease severity. This makes eye screenings more accessible and efficient, especially in areas with limited medical resources.
Eye diseases such as age-related macular degeneration (AMD), glaucoma, diabetic retinopathy (DR), diabetic macular edema (DME), and pathological myopia affect hundreds of millions of people globally. Early detection and screening are vital for managing these conditions and preventing vision loss. However, access to medical care, especially in low- and middle-income countries, remains a significant challenge, leading to a high burden of blindness worldwide.
To address this critical need, researchers have developed InSight, an innovative AI-based mobile application. InSight aims to make eye disease screenings more accessible by combining patient health information (metadata) with images of the eye’s retina (fundus images) to accurately diagnose five common eye diseases simultaneously on a mobile phone.
How InSight Works: A Three-Stage Approach
The InSight system operates through a clever three-stage pipeline designed for efficiency and accuracy. First, an image quality checker ensures that only clear, usable fundus images are processed. If an image is blurry or has poor lighting, the app prompts the user to retake it, achieving nearly 100% accuracy in filtering out low-quality images.
The second and core component is the disease diagnosis model. This model incorporates several key innovations. It uses a technique called MetaFusion to combine clinical metadata, such as a patient’s age, gender, diabetes diagnosis, and duration of diabetes, with the fundus images. This multimodal fusion allows the AI to consider a broader range of information, leading to more accurate diagnoses. The model also benefits from a specialized pretraining method that leverages both supervised and self-supervised learning, enabling it to learn effectively even with limited medical data. Furthermore, InSight employs a multitask model, meaning it can predict all five diseases at once. This approach is highly efficient, being five times more computationally efficient than using five separate models for each disease, making it ideal for deployment on mobile devices.
Finally, if diabetic retinopathy (DR) is detected, InSight includes a DR grading model. This third stage assesses the severity of DR, distinguishing between mild nonproliferative DR and severe proliferative DR, providing more detailed information beyond a simple yes/no diagnosis.
Promising Results and Real-World Applicability
InSight has demonstrated impressive performance. The complete pipeline consistently outperforms traditional image-only models, showing an average improvement of 11.2% in balanced accuracy across the five diseases. The multimodal disease diagnosis model, specifically, showed a 6% improvement in balanced accuracy on lab-captured images and a 4% improvement on smartphone-captured images compared to models that only use images.
The system achieved high diagnostic accuracy, with an Area Under the Receiver Operating Characteristic (AUROC) of 0.99 for diabetic retinopathy, pathological myopia, and AMD, 0.98 for diabetic macular edema, and 0.92 for glaucoma on the BRSET dataset. Importantly, InSight proved robust and generalized well to images captured by both traditional lab equipment and smartphone cameras, highlighting its potential for real-world use in diverse settings.
Understanding the Model’s Decisions
To build trust in its predictions, InSight is designed to be interpretable. Analysis showed that the model focuses on the correct anatomical regions and indicators within the fundus images for each disease. For instance, for diabetic retinopathy, the model correctly identifies regions likely to contain hemorrhages. The integration of metadata also proved crucial; for example, a strong correlation was observed between a patient’s diabetes status and the diagnosis of diabetic retinopathy and diabetic macular edema, which aligns with medical understanding.
Also Read:
- Decoding Radiologists’ Gaze: A New AI Model Uncovers Diagnostic Intent
- Advancing Newborn Pulmonary Hypertension Diagnosis with Multi-View AI
Looking Ahead
While InSight represents a significant step forward, the researchers acknowledge certain limitations. The study primarily used data from patients in Brazil, which might affect its generalizability to other populations. Additionally, some patient data was incomplete, requiring imputation techniques. Future work will involve studying InSight’s implementation in real-world screening programs and potentially expanding its capabilities to screen for even more eye diseases. For more in-depth information, you can read the full research paper here.
In conclusion, InSight establishes a scalable and accessible foundation for AI-assisted eye disease screening. By moving beyond single-disease models, integrating multimodal data, and optimizing for lightweight mobile platforms, this approach offers a promising solution for improving early diagnosis and access to eye care globally.
