How whole slide image registration supports AI in pathology

At Bigpicture, researchers are integrating a highly optimized WSI registration algorithm into the platform to improve pathology workflows and AI readiness. Daniel Budelmann, scientist at Fraunhofer Institute for Digital Medicine MEVIS, along with his colleagues Johannes Lotz and Nick Weiss, has been working on this challenge, adapting an algorithm originally developed for radiology and applying it to histology and pathology within Bigpicture.

"Whole Slide Image registration is essential for aligning different stained images of the same tissue sample," says Daniel. "Pathologists often review an H&E-stained slide and later request additional stainings, like IHC, to gain more information. But when they receive the new slide days later, they need to manually find the same area again. Registration solves this problem by automatically aligning the images, making the diagnostic process more efficient."

The process of a proven algorithm

The registration algorithm Budelmann’s team is implementing in Bigpicture was originally developed for radiology and radiotherapy, where image alignment is also critical. After proving successful in that field, the method was extended to histology and pathology, demonstrating fast, automatic, and highly accurate registration. "The core algorithm has been stable for years, and unlike many modern tools, it does not rely on AI," Daniel explains. "That means we don’t need to retrain it for different datasets, making it more robust across different types of stainings."

The process follows three main steps:

1. Pre-alignment: Roughly aligns images by correcting rotations and shifts.
2. Parametric Registration: Further refines the alignment using transformations like scaling, shear, and rotation.
3. Non-Parametric Registration: Fine-tunes local deformations pixel by pixel for high-precision alignment.

These steps allow researchers and AI models to precisely compare multi-stain images, enabling better biomarker discovery, AI training, and clinical diagnostics.

Demonstration of the WSI Registration Tool

Integration and optimization for Bigpicture users

As part of Bigpicture’s innovation efforts, the WSI registration algorithm is being integrated into the platform, making it available to beneficiary researchers and pathology labs.

"Originally, the implementation allowed users to select two images and register them," Daniel says. "However, the process of applying deformations to entire WSIs took a long time, up to 30 minutes per image. With recent improvements in the platform, we’re optimizing the process to enable real-time deformations on smaller image tiles instead of full slides." This optimization will allow users to apply fast, interactive image registration directly within Bigpicture, making it easier to analyze multi-stain tissue samples without external software.

Why this matters for digital pathology

Accurate alignment of WSIs has far-reaching implications across AI development, pathology research, and clinical diagnostics. By aligning different stainings, researchers can uncover new biomarkers and generate ground truth data essential for training AI models, such as those that can later identify tissue structures using just H&E staining. In clinical workflows, it reduces the time and effort needed to manually match areas between slides, making diagnostics faster and more reliable. “If we don’t standardize how slides are aligned, AI models trained on multi-stain datasets may struggle with inconsistencies,” says Daniel. “By integrating this tool into Bigpicture, we’re ensuring that AI applications in pathology have a strong foundation of high-quality, aligned data.”

Next steps: roll-out and further testing

The WSI registration tool is currently being integrated into Bigpicture’s Cytomine-based platform, with efforts focused on performance testing, user interface refinement, and ensuring compatibility across diverse datasets. A Python package is already available for partners who wish to begin testing the tool independently. “The more labs that use it, the better we can refine it for real-world pathology applications,” says Daniel. Full implementation is expected later this year, and Bigpicture partners are encouraged to get involved, share feedback, and help shape this essential tool for digital and AI-driven pathology.

About Daniel

Daniel Budelmann is a Scientist at Fraunhofer MEVIS, with over six years of experience in medical computer science since earning his MSc from the University of Lübeck. He specializes in image registration across radiology, radiotherapy, and histology, working on GPU algorithms to improve performance. He is also involved in the QuaLiPerF project, focusing on quantifying physiological processes in the liver. Proficient in software engineering with skills in C++, CUDA, and Python, he is dedicated to translating innovative technologies into practical applications.