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

Elena Buglakova, Anwai Archit, Edoardo D'Imprima, Julia Mahamid, Constantin Pape, Anna KreshukICCV 2025

TL;DR

This paper tackles the problem of "tiling artifacts" in the segmentation of large biological images, which arise from processing images in smaller chunks. They identify feature normalization as the root cause and propose Batch Renormalization as a solution that achieves seamless stitching without sacrificing prediction quality, a common trade-off with existing methods.

What Problem Does This Paper Address?

Segmenting large biological images, especially from microscopy, is crucial for scientific discovery but poses a significant computational challenge. These images often exceed GPU memory, necessitating a "sliding window" approach where the image is divided into smaller tiles for processing. While this allows for inference, it can lead to "tiling artifacts" – visual discontinuities or incorrect predictions at the boundaries between tiles. These artifacts can manifest as abrupt changes in segmentation or even "hallucinations" in background regions. Existing methods often rely on post-processing to mitigate these issues, making pipelines more complex and less reusable. This paper investigates the underlying cause of these artifacts, which they pinpoint to the feature normalization layers within neural networks, and explores the trade-offs between artifact-free segmentation and prediction quality.

Methods

The authors propose a systematic investigation into the causes of tiling artifacts. Their core methodology involves:

  • Analyzing Normalization Strategies: They compare different normalization techniques (BatchNorm, InstanceNorm, BatchRenorm, Identity) within standard segmentation architectures (3D U-Net and 3D UNETR).
  • Defining Metrics for Artifact Detection: To quantify the problem, they introduce two key metrics:
  • Train/Eval Disparity: Measures the difference in prediction quality between training and evaluation modes when using global normalization with running averages. This highlights performance drops not due to overfitting.
  • Tile Mismatch: Quantifies the discrepancy in predictions within the overlapping regions of adjacent tiles, directly measuring tiling artifacts.
  • Proposing Batch Renormalization (BatchRenorm): They advocate for BatchRenorm as a solution. This method uses global normalization statistics during both training and inference, effectively eliminating the train/eval disparity seen with BatchNorm and the tiling artifacts caused by InstanceNorm.

Datasets

The paper evaluates their methods on three diverse microscopy datasets:

  • Organoids [7]: Semantic segmentation of tissue in patient-derived colorectal cancer organoids (electron microscopy).
  • Plants [25]: Instance segmentation of nuclei in Arabidopsis ovules (confocal microscopy).
  • Embryo [2]: Instance segmentation of nuclei in mouse embryos (confocal microscopy).

For the "transfer setup," they also evaluate models trained on the embryo dataset for segmenting plants, and vice versa, to assess robustness.

Key Insights and Ideas

  • Normalization is the Culprit: The paper convincingly demonstrates that feature normalization layers, particularly tile-wise normalization like InstanceNorm, are the primary cause of tiling artifacts, not just edge effects.
  • The Normalization Trade-off: There's a fundamental trade-off: tile-wise normalization (InstanceNorm) leads to artifacts but can achieve good prediction quality, while global normalization (BatchNorm) avoids artifacts but can suffer from a significant performance drop during inference due to unstable batch statistics.
  • Batch Size Matters for BatchNorm: Small batch sizes, common in large image segmentation, make BatchNorm's running average statistics unreliable, leading to the train/eval disparity.
  • BatchRenorm as a Sweet Spot: Batch Renormalization offers a compelling solution by leveraging global statistics for seamless stitching without the performance penalty of BatchNorm or the artifacts of InstanceNorm.
  • Artifacts Worsen in Transfer Learning: Tiling artifacts become more pronounced when models are transferred to new datasets, highlighting the need for robust normalization strategies.

Key Results

  • InstanceNorm Causes Significant Tile Mismatch: Experiments show that InstanceNorm leads to noticeable tile mismatch artifacts across various datasets and architectures.
  • BatchNorm Suffers from Train/Eval Disparity: For datasets like organoids, BatchNorm exhibits high train/eval disparity (e.g., Dice score difference of 0.84), indicating a performance loss during inference.
  • BatchRenorm Achieves Artifact-Free and High-Quality Segmentation: BatchRenorm consistently produces results comparable to or better than InstanceNorm in terms of prediction quality (Dice scores) while completely eliminating tile mismatch artifacts. It also resolves the train/eval disparity issue.
  • Transfer Learning Benefits from BatchRenorm: The proposed BatchRenorm method shows improved performance and reduced artifacts in transfer learning scenarios.

Limitations and Future Work

  • Hyperparameter Tuning for BatchRenorm: While BatchRenorm is effective, the paper notes that finding the optimal switching point between BatchNorm and BatchRenorm during training might require some hyperparameter tuning.
  • Generalizability of Metrics: The proposed tile mismatch and train/eval disparity metrics are presented as general tools for diagnosing normalization issues in sliding window pipelines, suggesting their applicability beyond segmentation.
  • Further Exploration of Normalization: The paper focuses on specific normalization layers. Future work could explore other normalization techniques or hybrid approaches.
  • Dataset and Architecture Diversity: While tested on multiple datasets and architectures, exploring a wider range of biological image modalities and network designs could further validate the findings.

Summary updated

BibTeX
@inproceedings{buglakova2025tiling,
  author = {Elena Buglakova and Anwai Archit and Edoardo D'Imprima and Julia Mahamid and Constantin Pape and Anna Kreshuk},
  title = {{Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2025},
}
Figure