Self-Path: Self-Supervision for Classification of Pathology Images With Limited Annotations

Area of Science:

• Computational pathology
• Medical image analysis
• Deep learning

Background:

• High-resolution pathology images are valuable for deep learning but require extensive annotations.
• Acquiring comprehensive annotations for pathology images is a significant bottleneck for model training.

Purpose of the Study:

• To develop a self-supervised convolutional neural network (CNN) framework to learn generalizable representations from unlabeled pathology images.
• To improve semi-supervised learning and domain adaptation in computational pathology using limited labeled data.

Main Methods:

• Proposed a multi-task learning framework (Self-Path) using a CNN for tissue classification.
• Incorporated pathology-specific self-supervised pretext tasks leveraging contextual, multi-resolution, and semantic features.
• Evaluated the framework on three distinct pathology datasets.

Main Results:

• Self-Path achieved state-of-the-art performance in semi-supervised learning with minimal labeled data.
• Demonstrated improved domain adaptation for histopathology image classification without target domain labels.
• Showcased the framework's effectiveness across different pathology datasets.

Conclusions:

• Self-Path effectively utilizes unlabeled data for learning robust representations in pathology.
• The proposed pathology-specific self-supervision tasks enhance performance in low-data regimes and facilitate domain adaptation.
• This approach offers a viable solution for computational pathology applications with limited annotation budgets.