Classify epithelium-stroma in histopathological images based on deep transferable network.

Area of Science:

• Medical image analysis
• Computational pathology
• Artificial intelligence in medicine

Background:

• Deep learning methods are increasingly used in histopathological image analysis.
• Traditional deep learning models require data with identical distributions, limiting real-world applications.
• Acquiring large labeled datasets for each new imaging procedure is costly and time-consuming.

Purpose of the Study:

• To introduce an unsupervised domain adaptation technique for deep convolutional neural network (CNN) models in histopathology.
• To mitigate the need for extensive relabeling of data across different histopathological image acquisition procedures.
• To enhance the performance and applicability of CNNs in diverse real-world histopathological scenarios.

Main Methods:

• Implemented unsupervised domain adaptation by integrating feature-based adaptation and entropy minimization regularization terms.
• Modified a widely used CNN model, AlexNet, by incorporating these regularization terms into its objective function.
• Validated the proposed method using three independent public epithelium-stroma datasets.

Main Results:

• The proposed unsupervised domain adaptation method demonstrated superior performance in epithelium-stroma classification compared to traditional deep learning approaches.
• The method outperformed existing deep domain adaptation techniques on the tested datasets.
• Experimental results confirmed the effectiveness of the approach in handling domain shifts in histopathological data.

Conclusions:

• The developed unsupervised domain adaptation method offers a viable solution for histopathological image analysis.
• It effectively addresses the challenge of domain shift without requiring large-scale labeled data collection for new domains.
• This approach represents a significant advancement for practical, real-world applications in computational pathology.