AI-Driven Stitching of Retinal Smartphone-Based Images for Automated Fundoscopy

In this work, a lightweight convolutional neural network (CNN) and homography estimation are used to stitch low-resolution retinal images taken with smartphone devices. A Siamese MobileNetV2 network is used for feature extraction, and a simplified CNN-based regression model is used to generate the homography matrix, which allows for accurate image alignment by predicting corner point displacements. The model achieved remarkable performance on training with the APTOS, FIRE, and Retina Bank datasets with PSNR of 37.29 dB, SSIM of 0.88, and RMSE of 3.54. These results maintain the computational efficiency required for real-time mobile applications while outperforming competing methods. The suggested method guarantees high-accuracy image registration, ensuring more dependable automated retinal screenings, commanding in identifying conditions like glaucoma and diabetic retinopathy.