Evaluation of a Task-Specific Self-Supervised Learning Framework in Digital Pathology Relative to Transfer Learning...

Area of Science:

• Digital Pathology
• Computer Vision
• Machine Learning

Background:

• Digital pathology workflows rely on extracting features from whole-slide images.
• ImageNet-pretrained neural networks are commonly used for tile encoding.
• Evaluating different weight initialization strategies is crucial for optimizing performance.

Purpose of the Study:

• To critically analyze various tile encoding strategies in digital pathology.
• To identify the most effective approach for transfer learning in histopathology.
• To propose and evaluate a novel task-specific self-supervised learning framework.

Main Methods:

• Categorized neural network performance based on random, ImageNet-based, and self-supervised learning weight initialization.
• Developed a framework using task-specific self-supervised learning with a shallow feature extractor and spatial-channel attention block.
• Evaluated the framework on patch classification and weakly supervised whole-slide image classification tasks across multiple datasets.

Main Results:

• The proposed task-specific self-supervised encoding approach consistently outperformed other convolutional neural network-based encoders.
• Demonstrated superior performance across diverse datasets including colorectal cancer, Patch Camelyon, and prostate cancer detection.
• Highlighted the effectiveness of task-specific attention-based self-supervised training for histopathology.

Conclusions:

• Task-specific self-supervised learning enables tailored feature extraction for histopathology.
• This approach offers a promising alternative to using pretrained models from outside the histopathology domain.
• Supports the development of more focused and domain-specific analyses in digital pathology.