Patch Clustering for Representation of Histopathology Images

Wafa Chenni, Habib Herbi, Morteza Babaie, Hamid R. TizhooshECDP 2019 · DOI

Abstract

This paper presents an unsupervised pipeline to address the computational challenges of processing high-resolution Whole Slide Images (WSIs) in digital pathology. By utilizing Self-Organizing Maps (SOM) for clustering and Gaussian Mixture Models (GMM) for patch selection, the proposed method represents WSIs using only 10% to 50% of their constituent patches. Results on the KimiaPath24 dataset demonstrate that this compact representation maintains high retrieval accuracy while reducing storage and search time by up to 90%.

Introduction

In digital pathology, a single Whole Slide Image (WSI) can exceed 100,000 by 100,000 pixels. These gargantuan files create massive hurdles for real-time processing, memory usage, and storage. While Content-Based Image Retrieval (CBIR)—finding similar historical cases to assist in diagnosis—is a promising tool, searching through millions of high-resolution patches is computationally expensive.

Furthermore, most medical image data is unlabeled, and manual delineation of regions of interest is a costly task. This research seeks to create a "compact representation" of these slides that is memory-efficient and computationally inexpensive without requiring human-labeled data.

Methodology

The authors propose a multi-stage unsupervised pipeline to distill a WSI down to its most representative parts:

  1. Preprocessing: WSIs are broken into patches, and background patches (those with >99% white space) are discarded.
  2. Feature Extraction: The study compared handcrafted features (Local Binary Patterns, LBP) against deep features extracted from a pre-trained VGG16 network (reduced via PCA).
  3. Clustering (SOM): A Self-Organizing Map groups similar patches together, identifying different tissue types or textures without supervision.
  4. Patch Selection (GMM): A Gaussian Mixture Model selects a subset (10%–50%) of the most representative patches from each cluster.
Diverse tissue types from the KimiaPath24 dataset used in this study.

Experimental Results

The study utilized the KimiaPath24 dataset, which includes 24 WSIs representing diverse organs and tissue types (as shown in the figure above). The training set consisted of 27,055 patches after background removal.

Clustering Performance

The effectiveness of the SOM clustering was evaluated using the ratio of inter-cluster to intra-cluster distances. A higher ratio indicates better separation between different tissue textures. As shown in the figure below, increasing the SOM map size generally improves the distinctness of the clusters, allowing for more precise patch selection.

Ratio of inter-cluster to intra-cluster distances across different SOM map sizes, indicating clustering quality.

Retrieval Accuracy and Efficiency

  • Data Reduction: Using 100% of the patches yielded a retrieval accuracy of 69%. By using only 50% of the patches, the accuracy only dropped slightly to 65%.
  • Efficiency: The method achieved a reduction in search time and storage requirements of 50% to 90%.
  • Feature Comparison: Surprisingly, handcrafted LBP features often outperformed deep VGG16 features. LBP is exceptionally good at capturing the specific repetitive textures found in histopathology, whereas VGG16 (trained on natural images) may produce overly complex feature vectors for this specific task.

Discussion and Conclusion

The "less is more" philosophy proves effective here; by intelligently selecting 50% of the patches, the system retains enough information to perform nearly as well as the full dataset. The combination of SOM and GMM acts as a "de-duplicator," recognizing redundancy in uniform healthy tissue and saving the data "budget" for unique structures.

However, the authors acknowledge a risk: while excellent at recognizing dominant patterns, the system might miss minute cellular details or small, isolated tumors if they are not selected as "representative." Future work will explore specialized neural networks and refined selection strategies to ensure rare but critical features, such as micro-metastases, are not discarded.

Summary updated

BibTeX
@inproceedings{chenni2019patch,
  author = {Wafa Chenni and Habib Herbi and Morteza Babaie and Hamid R. Tizhoosh},
  title = {{Patch Clustering for Representation of Histopathology Images}},
  booktitle = {ECDP},
  year = {2019},
  doi = {10.1007/978-3-030-23937-4\_4},
}
Figure