Deep embeddings and logistic regression for rapid active learning in histopathological images

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Accurate tissue component recognition is crucial in digital pathology.
• Current Convolutional Neural Network (CNN) methods require extensive annotated data, which is labor-intensive to create.
• Interactive data annotation presents a potential solution to streamline dataset creation.

Purpose of the Study:

• To introduce Deep Embedding-based Logistic Regression (DELR) for efficient histopathological image analysis.
• To enable rapid model training and inference in digital pathology workflows.
• To explore DELR's effectiveness within an active learning framework.

Main Methods:

• DELR employs a pretrained CNN to generate compact image embeddings at low computational cost.
• A Logistic Regression model is trained efficiently using these embeddings.
• The method was implemented in an active learning framework and validated on lung, breast, and colorectal cancer datasets.

Main Results:

• DELR achieved an Area Under Curve (AUC) > 0.95 with only 100 image patches per class across three datasets.
• While slightly less accurate than fine-tuned CNNs, DELR demonstrated significant speed improvements (536x, 316x, 1481x faster post-encoding).
• DELR proved compatible with various active learning strategies and encoders.

Conclusions:

• DELR provides comparable accuracy to CNNs with substantially faster processing speeds.
• Its efficiency makes DELR a promising solution for real-time interactive data annotation in digital pathology.
• This method addresses the bottleneck of large-scale dataset creation for histopathological image analysis.