A Bi-channel Aided Stitching of Atomic Force Microscopy Images

Huanhuan Zhao, Ruben Millan-Solsona, Checa Nualart Marti, Spenser R. Brown, Jennifer Lynn Morrell-Falvey, Liam Collins, Arpan BiswasarXiv 2025 · DOI · arXiv

Introduction

Atomic Force Microscopy (AFM) is a critical tool for high-resolution imaging of biological samples, such as bacterial biofilms. However, researchers often face a fundamental trade-off between spatial resolution and the field-of-view (FOV). While high-magnification scans reveal fine structural details, they restrict observation to a minute fraction of the sample. To reconstruct larger areas, a grid of overlapping images is typically acquired and "stitched" together.

Traditional stitching methods rely on identifying matching landmarks across overlapping regions. Unfortunately, AFM topographical (height) maps are frequently "feature-sparse," characterized by smooth gradients that lack distinct landmarks for reliable alignment. This problem is exacerbated when image overlap is minimized (e.g., 10%) to reduce acquisition time and cost. In such scenarios, conventional tools often fail or produce significant artifacts, such as "seam lines," which can compromise scientific interpretation.

Methodology

To address these challenges, we propose a bi-channel aided feature-based stitching workflow. This method leverages secondary, feature-rich data channels collected simultaneously with the topographical map, such as the amplitude channel.

Workflow

The proposed pipeline consists of three primary stages:

  1. Feature Detection and Matching: The Scale-Invariant Feature Transform (SIFT) algorithm is applied to the amplitude channel. Unlike height maps, amplitude images capture the AFM tip's vibrational response to surface obstacles, naturally highlighting edges and fine textures.
  2. Camera Pose Estimation: High-quality feature matches from the amplitude channel are used to calculate the precise rotation, displacement, and tilt (camera pose) required for alignment.
  3. Warping and Blending: The calculated transformations are applied to the primary topographical images. A blending algorithm then merges the warped images into a seamless composite.
Comparison of traditional topographical stitching (a) and the proposed bi-channel aided stitching (b). The bi-channel approach successfully aligns images even at low overlap.

Synthetic Feature Channels

For datasets where secondary channels were not recorded, the method can be generalized. We found that the x-axis derivative of the topographical image serves as an effective synthetic proxy, creating a feature-rich map that mimics the edge-detection capabilities of the amplitude channel.

Results

The efficacy of the bi-channel method was evaluated using datasets of Pantoea sp. YR343 bacterial biofilms.

  • Feature Density: In a 3x3 grid with 10% overlap (Dataset 1), the topographical channel yielded only ~375 features. In contrast, the amplitude channel provided 2,746 features—a nearly seven-fold increase in data points for alignment.
  • Stitching Performance: At 10% overlap, traditional tools like Stitch2D failed to find a solution, while MIST produced composites with visible misalignments. The bi-channel method successfully generated seamless reconstructions.
  • Accuracy Validation: On feature-rich samples where direct stitching was possible (Dataset 3), the bi-channel method achieved a Structural Similarity Index (SSIM) of 0.99, confirming high geometric and radiometric accuracy.

Discussion and Conclusion

The bi-channel approach significantly improves the robustness of AFM image stitching by exploiting the "edge advantage" of amplitude data. By enabling accurate reconstruction at just 10% overlap, the method allows for more efficient large-area mapping without sacrificing resolution. Furthermore, the use of mathematical proxies like the x-derivative ensures the technique is applicable to legacy datasets.

Future work will focus on automating the selection of the optimal secondary channel and integrating deep learning-based matchers (e.g., SuperGlue) to enhance performance in noisy environments. Real-time implementation within the AFM control software could also provide researchers with immediate "big picture" feedback during scanning.

Summary updated

BibTeX
@misc{zhao2025bi,
  author = {Huanhuan Zhao and Ruben Millan-Solsona and Checa Nualart Marti and Spenser R. Brown and Jennifer Lynn Morrell-Falvey and Liam Collins and Arpan Biswas},
  title = {{A Bi-channel Aided Stitching of Atomic Force Microscopy Images}},
  year = {2025},
  eprint = {2503.08735},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.08735},
}
Figure