Abstract
<title>Abstract</title> <p>Recent advances in AI-based image analysis have substantially improved the efficiency of extracting meaningful information from large-scale image datasets. However, the effectiveness of these methods strongly depends on the quality of input data, necessitating robust and standardized image preprocessing steps to ensure accurate and reproducible results. In this study, we present a tailored preprocessing workflow designed to address the specific challenges associated with the registration of multimodal microscopy data. The core objective of this workflow is to prepare for the precise spatial alignment of images acquired from disparate imaging modalities, specifically electron microscopy (EM) and light microscopy (LM). Due to fundamental differences in contrast mechanisms, resolution, and field of view, direct correlation between such images remains a non-trivial task. Our approach aims to generate image pairs representing the same region of interest, thereby facilitating accurate registration. To this end, we implement a normalized cross-correlation-based template matching algorithm to localize corresponding features across modalities. In addition, to reconcile disparities in pixel dimensions, we exploit the presence of common structural landmarks in both image types, which serve as internal standard for calibrating the pixel size. Pixel size calibration constitutes a critical step in the preprocessing pipeline, as it directly impacts the performance of template matching and downstream registration procedures. Notably, this step currently requires manual supervision, as there is a lack of computational tools capable of autonomously adjusting pixel dimensions across heterogeneous imaging datasets. Upon completion of these preprocessing steps, the resulting image pairs are suitable for subsequent image analysis tasks, including registration, segmentation, and quantitative assessment. Moreover, the proposed weakly supervised preprocessing framework offers potential for integration into automated microscopy platforms, thereby enabling unsupervised, high-throughput processing of multimodal image data.</p>