Cross-Reference Attention for Region-Aware Super-Resolution of Blood Cell Images

Abstract Blood cell image analysis plays a critical role in clinical diagnostics, as white blood cells (WBCs) provide diagnostic cues for infections, cancers, and immune disorders. While low-magnification microscopy offers a wide field of view (FOV) but insufficient resolution for meaningful assessment, high-magnification microscopy offers more detailed but narrower FOV images. To obtain a high-magnification, large FOV image, current practice relies on stitching multiple high-magnification images, which is labor-intensive and inefficient. To address this, we propose a multi-reference-based super-resolution (RefSR) framework specifically designed for blood cell microscopy. By using paired low- and high-magnification images directly in the dataset construction process, we ensure robustness in real-world microscopy. Building on this, we develop a region-aware architecture that selectively applies RefSR to WBC regions and single-image SR (SISR) to background regions, thereby maintaining performance while reducing computational cost. Furthermore, we introduce a Cross-Reference Attention Module (CRAM) that explicitly models interactions between multiple high-magnification references, which conventional multi-RefSR methods ignore. Experiments show that our framework creates large FOV images with sharp WBC morphology and fewer artifacts than existing SR approaches, while achieving higher efficiency than the stitching method. These results demonstrate that our method provides a practical solution for accurate and scalable blood cell image analysis, effectively eliminating the need for conventional stitching. Graphical abstract Figure Graphical abstract showing the transformation from conventional image stitching approaches to the proposed super-resolution-based framework.