Efficient annotation bootstrapping for cell identification in follicular lymphoma

Area of Science:

• Digital pathology
• Computer vision
• Deep learning

Background:

• Digital pathology tasks heavily rely on visual cell and tissue pattern assessment.
• Acquiring sufficient annotations for deep learning is a major bottleneck due to cost, time, and label imbalance.
• This study addresses annotation cost reduction for follicular lymphoma diagnosis.

Purpose of the Study:

• To explore and compare annotation bootstrapping solutions for digital pathology.
• To reduce the cost and improve efficiency of data annotation for deep learning models.
• To enhance the detection of follicular lymphoma cells.

Main Methods:

• Comparison of three annotation bootstrapping approaches: manual, active learning, and weak supervision.
• Development of a hybrid architecture for cell detection using custom and foundation models.
• Utilized a dataset of 41 whole slide images with 12,704 cell annotations.

Main Results:

• Active learning workflow doubled minority class samples.
• The best bootstrapping method improved detection algorithm performance by 18 percentage points.
• Achieved a macro-averaged F1-score, precision, and recall of 63%.

Conclusions:

• Annotation bootstrapping, particularly active learning, is effective for imbalanced datasets in digital pathology.
• The proposed methods can improve deep learning model performance for follicular lymphoma diagnosis.
• Findings are applicable to other digital pathology tasks lacking homogeneous cell clusters.