IoT-Opthom-CAD: IoT-Enabled Classification System of Multiclass Retinal Eye Diseases Using Dynamic Swin Transformers and Explainable Artificial Intelligence

Abstract

Integrating Internet of Things (IoT)-assisted eye-related recognition incorporates connected devices and sensors for primary analysis and monitoring of eye conditions. Recent advancements in IoT-based retinal fundus recognition utilizing deep learning (DL) have significantly enhanced early analysis and monitoring of eye-related diseases. Ophthalmologists use retinal images in the diagnosis of different eye diseases. Numerous computer-aided diagnosis (CAD) studies have been conducted by using IoT and DL technologies on the early diagnosis of eye-related diseases. The retina is susceptible to microvascular alterations due to numerous retinal disorders. This study creates a new, non-invasive CAD system called IoTOpthom-CAD. It uses Swin transformers and the gradient boosting (LightGBM) method to find different eye diseases in colored fundus images after applying data augmentations techniques. We introduce a Swin transformer (dc-swin) that is efficient and powerful by connecting a dynamic cross-attention layer to extract local and global features. In practice, this dynamic attention layer suggests a mechanism where the model dynamically focuses on different parts of the image at other times, learning to cross-reference or integrate information across these parts. Next, the LightGBM method is used to divide these features into multiple groups, including normal (NML), diabetic retinopathy (DR), tessellation (TSN), age-related macular degeneration (ARMD), Optic Disc Edema (ODE), and hypertensive retinopathy (HR). To find the causes of eye-related diseases, the Grad-CAM is used as an explainable artificial intelligence (xAI). To develop the Opthom-CAD system, preprocessing, and data augmentation steps are integrated to strengthen this architecture. Multi-label three retinal disease datasets, such as MuReD, BRSET, and OIA-ODIR, are utilized to evaluate this system. After ten times of cross-validation tests, the proposed Opthom-CAD system shows excellent results such as an AUC of 0.95, f1-score of 95.7, accuracy of up to 96.5%, precision of 95%, recall of 94% and f1-score of 95.7. The results indicated that the performance of the Opthom-CAD system is much better than that of numerous baseline state-of-the-art models. As a result, the Opthom-CAD system can assist dermatologists in detecting eye-related diseases. The source code is public and accessible for anyone to view and modify from GitHub (https://github.com/Qaisar256/Opthom-CAD).