Robust registration under large image misalignment using an iterative step-aware transformer with application to corneal confocal microscopy

Corneal confocal microscopy (CCM) captures high-resolution images of the subbasal nerve plexus (SNP), but its small field of view limits visualization of larger anatomical landmarks such as the corneal whorl, which may aid in disease differentiation. Intra-modality image frame registration is, therefore, a critical step toward high-quality mosaic construction. However, raster scanning required for mosaicking in CCM is still performed manually, often leading to irregular coverage and substantial frame misalignment. This creates a challenge as existing deep learning methods largely focus on local elastic deformation, which is less relevant in CCM due to contemporaneous acquisition and local linearity, while paying limited attention to robust affine registration under large misalignment. To address this, we propose a hybrid cross- and self-attention transformer-based registration network that explicitly models long-range spatial correspondences, enabling robust affine alignment even under challenging conditions. To further enhance robustness, we introduce the Iterative Step-Aware Transformer Registration Network (ISATR-Net), which progressively reduces residual misalignment through step-aware conditioning. To rigorously evaluate this problem under the clinical diversity encountered in real-world CCM imaging, we curated a new CCM dataset. The dataset includes 156 eyes from 91 patients, spanning dry eye disease, neuropathic corneal pain, penetrating keratoplasty, uveitis, Fuchs corneal dystrophy, and control cases, totaling 1,349 image frame registration pairs. On this dataset, ISATR-Net outperforms multiple state-of-the-art affine registration methods. Our open-source codes are available online.