Unsupervised confocal superficial eyelid image stitching: Flexible, accurate and smooth

Abstract

Confocal laser scanning microscope enables non-invasive ocular Demodex screening but faces field-of-view (FOV) limitations. Although image stitching can theoretically expand FOV, traditional methods only achieve approximately 60% success rate due to low illumination, weak textures and repetitive patterns. To address these challenges, we propose an unsupervised deep learning-based image stitching framework with dual-stage alignment and generative adversarial network (GAN)-based fusion. Our dual-stage alignment network combines homography matrix and Thin Plate Spline (TPS) transformations to accommodate tissue deformation during imaging, supported by a Non-Maximum Suppression Feature Displacement Layer that simultaneously considers both long-range and short-range dependencies, yielding more accurate results with reduced memory consumption. To achieve smooth and seamless image fusion, we employ a GAN framework where the generator is designed to produce fusion probability maps that eliminate noticeable blending seams and fusion artifacts. This is an innovative attempt to apply deep learning for precise image stitching in confocal superficial eyelid image, demonstrating 40% higher success rate than traditional methods. Quantitative evaluations show 13.37% and 3.25% improvements in mPSNR and mSSIM over the state-of-the-art model, with 11.29% and 3.76% reductions in NIQE and PIQE metrics.