A deep active learning framework for mitotic figure detection with minimal manual annotation and labelling.

Area of Science:

• Digital pathology
• Artificial intelligence in oncology
• Computational cancer research

Background:

• Accurate mitotic figure (MF) identification is critical for glioblastoma (GBM) diagnosis and grading.
• Manual MF counting in whole slide images (WSIs) is time-consuming and subject to interobserver variability.
• There is a need for efficient and accurate automated methods for MF detection in GBM WSIs.

Purpose of the Study:

• To develop and evaluate a deep active learning framework for automated MF detection and classification in GBM WSIs.
• To minimize human intervention and annotation time in the MF identification process.
• To improve the accuracy and efficiency of MF analysis for GBM diagnosis.

Main Methods:

• Utilized a dataset of GBM WSIs from The Cancer Genome Atlas (TCGA).
• Integrated convolutional neural networks (CNNs) with an active learning strategy for iterative model training.
• Employed expert review for ambiguous cases identified by the framework to refine the model.

Main Results:

• Achieved 81.75% precision and 82.48% recall for MF detection.
• Attained 84.1% accuracy for MF subclass classification.
• Significantly reduced annotation time by approximately 900 minutes across 66 WSIs, nearly halving the effort.

Conclusions:

• The deep active learning framework offers substantial improvements in efficiency and accuracy for MF detection and classification in GBM WSIs.
• The approach reduces the need for large annotated datasets, minimizing manual effort while maintaining high performance.
• This methodology is generalizable to other medical imaging tasks, with potential broad applications in healthcare.