Improving acute lymphoblastic leukemia diagnosis through CBAM-enhanced VGG19 deep learning

Area of Science:

• Computational pathology
• Artificial intelligence in hematology
• Medical image analysis

Background:

• Manual review of bone marrow smears for Acute Lymphoblastic Leukemia (ALL) diagnosis is labor-intensive and subjective.
• Automated methods are crucial for improving the speed and accuracy of ALL detection and subtyping.

Purpose of the Study:

• To develop and evaluate a deep learning framework for automated detection and subtyping of ALL from microscopic bone marrow images.
• To enhance feature extraction and classification accuracy using attention mechanisms.

Main Methods:

• A hybrid CBAM-VGG19 deep learning model was developed, integrating a Convolutional Block Attention Module (CBAM) with a VGG19 backbone.
• The model hierarchically enhances morphological features in bone marrow images for improved ALL detection.
• K-fold cross-validation was employed to validate the model's performance.

Main Results:

• The CBAM-VGG19 model achieved a classification accuracy of 98.73%, outperforming other leading deep learning architectures.
• Attention mechanisms significantly improved feature extraction, learning speed, and classification accuracy, especially for similar leukemia subtypes.
• Image resolution, hyperparameter optimization, and CBAM layer placement analyses further enhanced model convergence and robustness.

Conclusions:

• The developed deep learning framework demonstrates high accuracy in automated ALL detection and subtyping from bone marrow images.
• While promising, the model is a research prototype requiring external validation and larger datasets for clinical application.
• This work provides a foundation for future large-scale, multi-center studies in AI-driven hematopathology.