Evaluation of two semi-supervised learning methods and their combination for automatic classification of bone marrow...

Area of Science:

• Hematology
• Medical Imaging
• Artificial Intelligence

Background:

• Differential bone marrow (BM) cell counting is crucial for diagnosing hematological diseases.
• Accurate BM cell classification is challenging due to non-uniformity and lack of reproducibility.
• Deep learning systems require large, labeled datasets for accurate BM cell classification.

Purpose of the Study:

• To develop an automatic BM cell classification system using semi-supervised learning (SSL).
• To evaluate the effectiveness of self-training (ST), active learning (AL), and a combined approach (CST+AL) for increasing labeled training data.
• To improve the accuracy of classifying 16 types of BM cell images.

Main Methods:

• Employed three SSL methods: confirmed self-training (CST), active learning (AL), and a combination (CST+AL).
• Iteratively expanded the training dataset over 25 rounds.
• Classified 16 types of BM cell images using deep learning models.

Main Results:

• The training dataset size increased from 425 to over 40,000 images with CST, 3,682 with AL, and 47,843 with CST+AL.
• Classification accuracies for test data were 0.944 (CST), 0.941 (AL), and 0.976 (CST+AL).
• The CST+AL method resulted in fewer class imbalances compared to CST or AL alone.

Conclusions:

• The combined CST+AL approach is efficient for augmenting training data in automatic BM cell classification systems.
• SSL methods, particularly the combined approach, significantly improve the accuracy and efficiency of BM cell classification.
• This study highlights the potential of advanced SSL techniques for developing robust diagnostic tools in hematology.