Deep Learning for Whole Slide Image Analysis: An Overview.

Area of Science:

• Digital pathology
• Computer vision
• Artificial intelligence

Background:

• Whole slide imaging (WSI) generates gigapixel-sized images, demanding advanced analysis techniques.
• Deep learning (DL) shows promise for visual understanding but struggles with WSI's scale, heterogeneity, and artifacts.
• Clinical translation of DL in digital pathology requires addressing these specific challenges.

Purpose of the Study:

• To review interdisciplinary approaches for training deep neural networks (DNNs) on WSI data.
• To highlight methodologies designed to overcome the limitations of applying DL to gigapixel pathology images.
• To provide insights into enabling DL for effective gigapixel image analysis in clinical settings.

Main Methods:

• Review of existing literature on DL applications in WSI analysis.
• Categorization of methodologies addressing WSI-specific challenges (e.g., memory constraints, artifacts, heterogeneity).
• Analysis of interdisciplinary strategies combining computer vision and pathology expertise.

Main Results:

• Identified key challenges in WSI analysis using DL: computational demands, data variability, and image artifacts.
• Highlighted various DL architectures and training strategies adapted for gigapixel images.
• Showcased successful adaptations of DL for tasks like tumor detection and classification on WSI.

Conclusions:

• Overcoming WSI-specific challenges is crucial for the clinical adoption of DL.
• Interdisciplinary research is vital for developing robust DL solutions for digital pathology.
• Future work should focus on standardization and validation of DL methods for WSI analysis.