Domain-Specific Pre-training Improves Confidence in Whole Slide Image Classification.

Area of Science:

• Digital pathology
• Computational pathology
• Machine learning in healthcare

Background:

• Whole Slide Images (WSIs) present significant challenges for deep learning due to their size and limited annotations.
• Multiple-instance learning (MIL) models are increasingly used for WSI analysis, often relying on generic pre-trained models like ResNet-50.
• Domain-specific pre-training, such as with KimiaNet (DenseNet121 pre-trained on TCGA slides), offers a specialized approach.

Purpose of the Study:

• To evaluate the impact of domain-specific pre-training on the performance of MIL models for WSI classification.
• To assess the effect of domain-specific pre-training on model confidence and predictive accuracy.
• To investigate the clinical applicability of these enhanced models in glioma diagnosis.

Main Methods:

• Utilized state-of-the-art MIL models: CLAM (attention-based) and TransMIL (self-attention-based).
• Compared models pre-trained with generic datasets versus domain-specific datasets (TCGA slides).
• Evaluated model confidence and predictive performance specifically for the detection and subtyping of gliomas from WSIs.

Main Results:

• Domain-specific pre-training significantly improved the confidence of both CLAM and TransMIL models.
• Achieved new state-of-the-art performance in WSI-based glioma subtype classification.
• Demonstrated enhanced predictive accuracy compared to models using generic pre-training.

Conclusions:

• Domain-specific pre-training is crucial for improving deep learning model performance in digital pathology.
• The enhanced models show high clinical applicability for assisting in glioma diagnosis and subtyping.
• Publicly shared code and results facilitate further research and development in computational pathology.