ActiveSSF: An Active-Learning-Guided Self-Supervised Framework for Long-Tailed Megakaryocyte Classification.

Area of Science:

• Medical Image Analysis
• Computational Pathology
• Hematology

Background:

• Accurate megakaryocyte classification is vital for diagnosing myelodysplastic syndromes.
• Self-supervised learning shows potential but faces challenges in medical imaging, including background noise, imbalanced data, and complex cell morphology.
• Existing methods struggle with the high intra-class variability and rare subtypes of megakaryocytes.

Purpose of the Study:

• To develop and validate the ActiveSSF framework for improved megakaryocyte classification.
• To address the limitations of self-supervised learning in analyzing complex medical images.
• To enhance diagnostic accuracy for myelodysplastic syndromes through precise megakaryocyte identification.

Main Methods:

• Proposed the ActiveSSF framework, integrating active learning with self-supervised pretraining.
• Utilized Gaussian filtering, K-means clustering, and HSV analysis with clinical prior knowledge for region-of-interest extraction.
• Implemented an adaptive sample selection mechanism to handle class imbalance and prototype clustering for morphological complexity.

Main Results:

• ActiveSSF achieved state-of-the-art performance on clinical megakaryocyte datasets.
• Demonstrated significant improvement in recognition accuracy for rare megakaryocyte subtypes.
• The framework effectively mitigated challenges posed by background noise, data imbalance, and morphological variations.

Conclusions:

• ActiveSSF offers a robust solution for megakaryocyte classification in challenging medical imaging scenarios.
• The framework holds significant practical potential for clinical applications in diagnosing myelodysplastic syndromes.
• Integration of active learning and self-supervised pretraining enhances diagnostic capabilities for hematological disorders.