Graph temporal ensembling based semi-supervised convolutional neural network with noisy labels for histopathology...

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in medicine

Background:

• Convolutional neural networks (CNNs) excel in histopathology image classification but demand extensive, accurate annotated data.
• Manual image labeling is time-consuming, costly, and prone to errors by pathologists.
• Existing methods struggle with noisy labels and the efficient utilization of unlabeled histopathology data.

Purpose of the Study:

• To develop a robust deep learning architecture for histopathology image classification that minimizes reliance on large-scale annotated datasets.
• To enhance the utilization of unlabeled data in the training process for improved classification performance.
• To create a model resilient to noisy labels commonly encountered in medical image annotation.

Main Methods:

• A novel self-ensembling deep architecture was proposed, integrating feature and label predictions.
• Exponential Moving Average (EMA) was used to aggregate predictions across epochs, creating ensemble targets.
• Clustering of ensemble targets within classes and a consistency cost were employed to enhance predictions and ensure consensus.

Main Results:

• The proposed method achieved 90.5% accuracy on lung cancer and 89.5% on breast cancer datasets using only 20% of labeled patients.
• Performance was comparable to baseline methods utilizing 100% of labeled data.
• The architecture demonstrated significant robustness against a small percentage of noisy labels.

Conclusions:

• The self-ensembling deep architecture effectively leverages limited labeled data and unlabeled data for histopathology image classification.
• The method offers a robust solution to the challenges of data annotation cost, time, and label noise in computational pathology.
• This approach shows potential for improving the efficiency and reliability of AI-driven diagnostic tools in cancer research.