Transferable automatic hematological cell classification: Overcoming data limitations with self-supervised learning.

Area of Science:

• Hematology
• Computational Biology
• Machine Learning

Background:

• Accurate classification of peripheral blood and bone marrow cells is crucial for diagnosing and monitoring hematological disorders.
• Current automatic classification systems face challenges due to data scarcity and limited generalizability across different laboratories.
• Self-supervised learning (SSL) offers a promising approach to overcome these limitations in cell classification.

Purpose of the Study:

• To integrate SSL into cell classification pipelines for hematological disorders.
• To address challenges of data scarcity and model generalizability in automated cell classification.
• To evaluate the performance of SSL-based feature extraction and classification compared to supervised methods.

Main Methods:

• Utilized four public hematological single-cell image datasets (one bone marrow, three peripheral blood).
• Employed an SSL-based approach for image feature extraction, requiring no initial image annotations.
• Applied a lightweight machine learning classifier trained on a small subset of annotated images using SSL features.

Main Results:

• SSL models trained on bone marrow data demonstrated higher classification accuracy when transferred to peripheral blood datasets compared to supervised deep learning models.
• After fine-tuning with 50 labeled samples per class, the SSL pipeline outperformed supervised deep learning for specific datasets and rare cell types.
• The SSL approach showed comparable performance to supervised methods on other datasets.

Conclusions:

• SSL enables the extraction of significant cell image features without relying on class labels.
• Knowledge transfer between bone marrow and peripheral blood cell domains is efficiently facilitated by SSL.
• SSL models demonstrate effective adaptation to new datasets with minimal labeled data, enhancing classification accuracy.