Engineered feature embeddings meet deep learning: A novel strategy to improve bone marrow cell classification and...

Area of Science:

• Hematology
• Medical Imaging
• Artificial Intelligence

Background:

• Manual cytomorphology evaluation of bone marrow cells is crucial for diagnosing hematological diseases but is time-consuming and can be a bottleneck.
• Current deep learning models for cell evaluation often focus on limited subtypes and lack interpretability.

Purpose of the Study:

• To develop an automated method for bone marrow cell evaluation using deep learning with enhanced interpretability.
• To improve the classification performance for a comprehensive set of 21 bone marrow cell subtypes.

Main Methods:

• Engineered a novel 'region-attention embedding' by organizing cytology features within a matrix, preserving spatial relations of segmented cell regions (cytoplasm, nucleus, whole-cell).
• Integrated this embedding with deep learning networks (Xception, ResNet50) to provide interpretable insights into image regions contributing to predictions.
• Evaluated the approach on a large public database (89,484 training, 22,371 testing images) with 21 cell subtypes using 3-fold cross-validation.

Main Results:

• The region-attention embedding combined with deep learning networks achieved a f1-score of 0.82 on the validation set, outperforming previous methods.
• The strategy demonstrated competitive performance on an unseen test set, achieving a f1-score of 0.56.
• The method provides interpretable information by highlighting relevant image regions.

Conclusions:

• The proposed region-attention embedding strategy significantly enhances deep learning-based bone marrow cell classification accuracy and interpretability.
• This approach offers a promising solution for automating hematological disease diagnosis, addressing the limitations of manual evaluation.