Reducing annotation effort in digital pathology: A Co-Representation learning framework for classification tasks.

Area of Science:

• Digital Pathology
• Computer Vision
• Machine Learning

Background:

• Digital pathology image classification is crucial for cancer diagnosis and prognosis.
• Deep learning offers automated solutions but requires large annotated datasets, which are costly and prone to variability.
• Existing methods face challenges in data acquisition and observer variability.

Purpose of the Study:

• To develop a deep learning framework for digital pathology image classification that minimizes the need for extensive annotated training data.
• To enhance the learning capability of deep neural networks in low-data regimes.
• To improve the accuracy and efficiency of automated pathology analysis.

Main Methods:

• Propose a co-representation learning framework combining categorical cross-entropy and deep metric learning objectives.
• Incorporate a deep metric learning objective to boost classification performance, particularly with limited data.
• Utilize a neighborhood-aware multiple similarity sampling strategy and a soft-multi-pair objective to accelerate deep metric learning.

Main Results:

• Achieved state-of-the-art performance on five benchmark datasets across nuclei classification, mitosis detection, and tissue type classification tasks using approximately 50% of training data.
• Outperformed existing state-of-the-art methods on all datasets when using complete training data.
• Demonstrated the framework's effectiveness in low-data scenarios.

Conclusions:

• The proposed co-representation learning framework effectively classifies digital pathology images with reduced training data requirements.
• The method offers a robust solution to the challenges of data acquisition and observer variability in digital pathology.
• This approach has the potential to significantly advance automated cancer diagnosis and prognosis tools.