Related Experiment Video
Updated: Jul 29, 2026
10:59
Reconstruction of 3-Dimensional Histology Volume and its Application to Study Mouse Mammary Glands
Published on: July 26, 2014
14.4K
Registered multi-device/staining histology image dataset for domain-agnostic machine learning models
Mieko Ochi1, Daisuke Komura2, Takumi Onoyama1,3
1Department of Preventive Medicine, Graduate School of Medicine, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo, 113-0033, Japan.
Scientific Data
|April 3, 2024
Summary
A new dataset, PathoLogy Images of Scanners and Mobile phones (PLISM), addresses biases in histopathology images caused by varied staining and imaging devices. This resource aids in developing more robust machine learning models for digital pathology.
Area of Science:
- Digital Pathology
- Computational Pathology
- Medical Imaging
Background:
- Histopathology image variations in color and texture arise from differing staining protocols and imaging devices across institutions.
- These domain shifts negatively impact the robustness and generalizability of machine learning models in digital pathology.
- Developing methods to mitigate these domain shifts is crucial for reliable automated analysis of histopathological data.
Purpose of the Study:
- To introduce the PathoLogy Images of Scanners and Mobile phones (PLISM) dataset, a comprehensive resource for evaluating domain shift in histopathology.
- To enable precise assessment of color and texture variations across diverse staining conditions and imaging devices.
- To facilitate the development and validation of robust machine learning models for histopathological image analysis.
Main Methods:
- The PLISM dataset was curated, comprising 46 human tissue types, 13 hematoxylin and eosin staining conditions, and 13 imaging devices.
- Precisely aligned image patches were generated to allow for quantitative evaluation of color and texture properties across different domains.
- A convolutional neural network pre-trained on the PLISM dataset was used to assess improvements in handling domain shift.
Main Results:
- Significant variations in color and texture were observed across different domains within the PLISM dataset, especially between whole-slide images and smartphone captures.
- The PLISM dataset demonstrated substantial diversity, reflecting real-world challenges in histopathology image acquisition.
- Pre-training a convolutional neural network on PLISM showed improvements in domain adaptation and robustness.
Conclusions:
- The PLISM dataset is a valuable resource for quantifying and addressing domain shift in digital pathology.
- This dataset supports the development of more resilient machine learning algorithms capable of handling variations in histopathology image data.
- PLISM contributes to advancing the reliability and applicability of AI in histopathological analysis across diverse clinical settings.
