Source-linked AI summary
The importance of stain normalization in colorectal tissue classification with convolutional networks
Francesco Ciompi, Oscar Geessink, Babak Ehteshami Bejnordi, Gabriel Silva de Souza, Alexi Baidoshvili, Geert Litjens, Bram van Ginneken, Iris Nagtegaal, Jeroen van der Laak
TL;DR
Reliable CRC imaging biomarkers require accurate and reproducible classification of tissue components in H&E histopathology images. The paper develops a ConvNet-based classifier, evaluates it on rectal-cancer and independent CRC data, and investigates stain normalization. The authors conclude that stain normalization is a necessary step in ConvNet-based CRC tissue-classification pipelines.
Problem
Manual and cross-source variability in H&E tissue appearance complicates reliable automated classification needed for CRC imaging biomarkers.
Method
The study trains and validates an 11-layer fully convolutional network on 74 rectal-cancer whole-slide images, then tests independent CRC images with two stain-normalization algorithms.
Results
Stain normalization improved ConvNet classification of independent CRC data, compared with the 50.96% accuracy obtained without normalization.
Takeaways & Limitations
The authors conclude that stain normalization is a necessary step in training and evaluation pipelines for ConvNet-based CRC tissue classification.
Abstract
from arXiv · showhide
The development of reliable imaging biomarkers for the analysis of colorectal cancer (CRC) in hematoxylin and eosin (H&E) stained histopathology images requires an accurate and reproducible classification of the main tissue components in the image. In this paper, we propose a system for CRC tissue classification based on convolutional networks (ConvNets). We investigate the importance of stain normalization in tissue classification of CRC tissue samples in H&E-stained images. Furthermore, we report the performance of ConvNets on a cohort of rectal cancer samples and on an independent publicly available dataset of colorectal H&E images.
1. INTRODUCTION
CRC biomarkers could improve patient stratification beyond tumor stage, but manual histological assessment is variable and difficult to apply clinically. The paper therefore examines ConvNets and stain normalization for reproducible automated tissue classification in H&E images.
- Clinical motivation: CRC has high incidence and mortality, while patients with the same tumor stage can have substantially different outcomes.Additional biomarkers are sought to identify patients who may or may not benefit from adjuvant chemotherapy.
- Clinical motivation: Histological parameters such as necrosis and tumor-to-stroma proportions are independent prognostic factors, but manual assessment has limited clinical applicability and high observer variability.These limitations motivate automated analysis of digitized whole-slide images.
- Automated pathology: Automated pixel-wise classification of tissues and histological structures is a key component of imaging-biomarker development from whole-slide images.ConvNets have become a reference approach for image- and patch-based medical-imaging classification, including H&E histopathology analysis.
- Stain variability: H&E appearance varies across laboratories and staining batches, and this variability can strongly affect automated image-analysis algorithms.Stain normalization algorithms aim to match whole-slide-image stain colors to a selected template.
- Study objective: The paper investigates stain normalization in ConvNet-based CRC tissue classification using a proposed 9-tissue-type rectal-cancer classifier and data from independent sources.It compares classification with and without two state-of-the-art stain-normalization algorithms and examines how normalization should be used in the pipeline.
2. METHOD
The method combines ConvNets for colorectal tissue classification with two stain-normalization algorithms and evaluates them across rectal-cancer and independent colorectal datasets.
- Data: The study uses 74 rectal-cancer whole-slide images with expert-checked annotations for nine tissue classes.The images were scanned at 200X, with 2µm H&E sections and 0.455µm pixel resolution.
- Data: An independent colorectal dataset contains 5,000 patches and ten large tiles spanning eight tissue types.The patches are 150×150 pixels, while the tiles are 5000×5000 pixels.
- ConvNet: The proposed model is an 11-layer fully convolutional network that maps 150×150 RGB patches to probabilities for nine RC tissue types.It alternates ReLU convolutional and max-pooling layers, doubling filters after each pooling layer.
- Stain normalization: Stain normalization transforms an image using parameters extracted from a template so its stain-color distribution resembles that template.The parameters capture color information from major stain components such as hematoxylin and eosin.
- Stain normalization: The evaluation compares two state-of-the-art algorithms: SN1, based on Bejnordi et al., and SN2, based on Macenko et al.Both methods are used to investigate stain normalization in colorectal tissue classification.
3. EXPERIMENTS
Experiments evaluate the ConvNet on rectal-cancer data and then test transfer to an independent colorectal dataset, including the effect of stain normalization in training and testing.
- RC evaluation: The RC evaluation uses 5-fold cross-validation with patient-disjoint training, validation, and test whole-slide images.Each fold selects 40 slides for training, 19 for validation, and 15 for testing.
- Independent CRC evaluation: Direct application to the independent CRC dataset achieved 50.96% accuracy in the six-class task.The RC and CRC class definitions were grouped into corresponding categories before evaluation.
- Independent CRC evaluation: 75.55% accuracy with SN1 and 73.99% with SN2 improved independent-CRC classification by more than 20% over direct application.The stain-normalized CRC patches were classified using the same ConvNet trained on RC data.
- Stain-normalization comparison: SN2 failed to normalize three adipose-tissue patches and tended to force color normalization regardless of tissue type.The reported qualitative comparison contrasts SN2 with SN1’s preservation of tissue color information.
- Stain-normalization comparison: SN2 completely removed the red color of blood cells, whereas SN1 retained a substantial red-channel component.This comparison is illustrated using CRC patches containing tumor and blood-cell tissue.
- Training and testing normalization: A further experiment evaluates SN1 applied to training data, test data, both, or neither in the RC-to-CRC pipeline.The four configurations are identified as experiments C and D alongside the preceding configurations.
4. CONCLUSION
The proposed ConvNet approach classifies CRC tissue in H&E histopathology images, and the experiments conclude that stain normalization is necessary in the training and evaluation pipeline.
- Stain normalization removes most staining-related variability from training and test data.
- The study concludes that stain normalization is necessary for automatic CRC tissue classification with ConvNets.