About AIRIN

Project Overview

The AIRIN project proposes an innovative solution that integrates advanced deep learning and explainable AI technologies to support medical decision-making in ophthalmology. Its primary goal is to develop a MDSS capable of analyzing ophthalmic images and providing accurate, explainable, and clinically actionable predictions.

AIRIN has been trained and validated on a broad spectrum of conditions — including choroidal neovascularization (CNV), drusen, macular hole (MH), central serous chorioretinopathy (CSR), and diabetic macular edema (DME)— as well as normal cases, thus covering a significant portion of clinical ophthalmic practice. The dataset, comprising both public and private sources, includes OCT images from over 18,000 patients, representing diverse centers and populations. Careful stratification by disease and center, combined with inter-expert agreement checks, reduced labelling variability and enhanced result reliability.

Strategy and Innovation

The strategy and innovation of AIRIN lie in a multidimensional approach that combines technological excellence, clinical validity, and educational value:

• Multiple diagnostic model selection: Users can choose between CNNs and LLMs, compare outputs, and strengthen decision-making. LLMs are trained on the same imaging data as CNNs and further enhanced with medical knowledge from textbooks, functioning as autonomous diagnostic models.
• Multi-XAI techniques: Integrating heatmaps (CAM) with textual explanations generated by LLMs, providing both visual and linguistic interpretability of predictions.
• Multi-external validation: Conducted across three external datasets (one from a public hospital and two from private ophthalmology centers in different cities) to ensure prediction stability, robustness, and generalizability.
• Ground truth protocol: A high-quality labelling protocol was applied to private datasets by four ophthalmologists of varying experience levels (resident >3 years, ophthalmologist >5 years, and two retina specialists >5 and >15 years), reducing inter-observer variability and creating a robust reference standard.
• Scalability: Initially trained on OCT images, AIRIN's architecture supports expansion to other imaging modalities (e.g., fundus photography, ocular ultrasound) and additional diseases, making it a multimodal and clinically adaptable diagnostic system.
• Educational dimension: Beyond diagnosis, the MDSS also serves as a training platform for medical students and ophthalmology residents.

Results

Normal/Abnormal Classification

The AIRIN system was initially trained and tested on the classification of OCT images as either normal or abnormal, and subsequently on the categorization of abnormal cases into five ophthalmic conditions (CNV, Drusen, MH, CSR, DME).

The high sensitivity translates into minimal false negatives, enabling patients with normal imaging to be safely reassured while allowing physicians to focus on complex cases. Clinically, this reduces waiting lists and unnecessary referrals.

Disease Classification Performance

For disease classification across the three external datasets, AIRIN achieved:

These high performances across independent external centers confirm the generalizability of the models across different populations, demonstrating the system's clinical reliability beyond controlled lab conditions.

Explainability and Clinical Integration

Among the explainability techniques, saliency maps, guided backpropagation, and Grad-CAM++ stood out and were evaluated by clinical experts as adequate. The LLMs enhance the diagnostic process by providing not only explanations but also autonomous predictions, combining OCT image analysis with knowledge from medical literature. The generated textual explanations are consistent with visual indicators (e.g., Grad-CAM), thereby increasing transparency and clinician trust. Moreover, the integration of Retrieval-Augmented Generation (RAG) allows the system to retrieve evidence-based information from medical texts in real time, enhancing both accuracy and educational value.

Benefits and Impact

At the application level, AIRIN delivers multiple benefits for healthcare providers: improving diagnostic accuracy and speed, facilitating faster patient flow, standardizing clinical reports, and reducing unnecessary tests and time costs, while socially promoting equitable access to quality care through smart pre-screening in resource-limited settings.

Conclusion

AIRIN integrates state-of-the-art artificial intelligence into a clinically critical specialty, offering a transparent, reliable, and scalable MDSS. It combines scientific innovation (multi-model diagnostic approach, multi-XAI, multi-external validation) with tangible social impact, bridging the gap between AI research and real-world clinical implementation. Its dual role as both a diagnostic and an educational tool amplifies its value—enhancing both the quality of care and the training of the next generation of physicians.