Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images

2025-03-25

AI Summary

This work identifies the root cause of stitching artifacts in large-scale biological image segmentation using sliding-window inference: statistical inconsistency across local windows induced by Batch Normalization (BN) layers. Through systematic normalization layer diagnostics, we propose the first quantitative metric for artifact detection and validate it on three real-world microscopy datasets. Results show BN is the dominant source of boundary artifacts, whereas BatchRenorm achieves the optimal trade-off between artifact suppression and segmentation accuracy. Furthermore, BatchRenorm significantly improves cross-dataset generalization and transfer performance—without requiring additional annotations or architectural modifications. Our approach provides a reproducible, plug-and-play normalization optimization framework specifically tailored for high-resolution medical and microscopic image segmentation.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Other Foundations of Machine Learning
• Reasoning under Uncertainty: Stochastic Optimization

Application Category

• Web Mining and Content Analysis: Normalization, clustering, classification, and summarization of Web text
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Security and Privacy: Data transparency and provenance

Abstract

Segmentation of very large images is a common problem in microscopy, medical imaging or remote sensing. The problem is usually addressed by sliding window inference, which can theoretically lead to seamlessly stitched predictions. However, in practice many of the popular pipelines still suffer from tiling artifacts. We investigate the root cause of these issues and show that they stem from the normalization layers within the neural networks. We propose indicators to detect normalization issues and further explore the trade-offs between artifact-free and high-quality predictions, using three diverse microscopy datasets as examples. Finally, we propose to use BatchRenorm as the most suitable normalization strategy, which effectively removes tiling artifacts and enhances transfer performance, thereby improving the reusability of trained networks for new datasets.

Problem

• Investigates tiling artifacts in large image segmentation
• Identifies normalization layers as root cause of artifacts
• Proposes BatchRenorm to eliminate artifacts and improve transfer

Innovation

• Investigates normalization layers causing tiling artifacts
• Proposes BatchRenorm for artifact-free segmentation
• Enhances transfer performance and network reusability