Scalable and Efficient Deep Learning-Based Pipeline for Mitotic Detection and Analysis in Pathology Images.

Area of Science:

• Digital pathology
• Computational oncology
• Image analysis

Background:

• Accurate mitotic figure analysis in whole-slide images (WSIs) is crucial for cancer grading and prognosis.
• Current methods face challenges in balancing accuracy with the high throughput required for clinical settings.

Purpose of the Study:

• To develop and validate a highly accurate and efficient three-stage pipeline for WSI-scale mitosis analysis.
• To refine mitosis detection, classify atypical mitoses, and assess prognostic value.

Main Methods:

• A three-stage pipeline utilizing YOLOv11 for candidate proposal, an ultra-lightweight classifier for refinement, and a downstream classifier for atypical mitosis identification.
• Validation on benchmark datasets and proof-of-concept survival analysis using TCGA-BRCA cohort data.

Main Results:

• The pipeline demonstrated improved F1 scores over detection-only methods and strong accuracy in atypical mitosis classification.
• Mitosis-derived features showed potential incremental prognostic value in early-stage breast cancer.
• Gigapixel WSIs were processed in minutes on a single GPU, indicating high efficiency.

Conclusions:

• The developed method offers accurate mitosis detection and robust classification of atypical forms with high efficiency.
• The pipeline is suitable for large-scale translational studies and potential integration into clinical workflows.
• Mitotic figure analysis holds promise for enhancing prognostic assessments in oncology.