A semi-supervised learning approach with consistency regularization for tumor histopathological images analysis.

Area of Science:

• Digital Pathology
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Manual histopathological image analysis is time-consuming and challenging for pathologists.
• Whole-slide imaging generates digital slices, necessitating advanced computer-aided diagnosis tools.
• Accurate classification of tumor subtypes is crucial for prognosis and treatment.

Purpose of the Study:

• To develop an automated system for classifying histopathological images into normal tissue and tumor subtypes.
• To address the challenge of limited labeled data in histopathology through semi-supervised learning.
• To enhance the flexibility and efficiency of histopathological analysis in clinical settings.

Main Methods:

• A semi-supervised learning algorithm, Semi-His-Net, was developed using a consistency regularization strategy.
• The model learns from both labeled and unlabeled data by enforcing consistent predictions on perturbed versions of the same image.
• This approach enables the model to generate pseudo-labels for unlabeled data, reducing the need for extensive manual annotation.

Main Results:

• Semi-His-Net achieved 90% accuracy in classifying breast cancer histopathological image patches into normal tissue and three tumor subtypes.
• The average Area Under the Curve (AUC) for cross-classification between tumor subtypes was 0.893.
• The model demonstrated effective classification performance, highlighting its potential in digital pathology.

Conclusions:

• Semi-His-Net offers a deep learning-based solution to overcome limitations of manual histopathology image review.
• The framework has the potential to significantly improve the efficiency and repeatability of cancer diagnosis.
• Automated classification of histopathological images can aid pathologists in diagnosis and prognostic evaluation.