CNN-Powered Dual-Path Network with Adaptive Attention for Red Blood Cell Classification.

Area of Science:

• Hematology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate red blood cell (RBC) morphology classification is vital for diagnosing hematological disorders.
• Human interpretation of RBC morphology is prone to errors due to subtle cellular differences.
• Deep learning and medical imaging offer promising avenues for automated RBC analysis.

Purpose of the Study:

• To introduce a novel deep learning architecture, CNN-DP-Att, for enhanced RBC morphology classification.
• To leverage dual-path networks and attention mechanisms for improved feature extraction.
• To achieve higher classification accuracy compared to existing methods.

Main Methods:

• Developed a dual-path convolutional neural network (CNN-DP-Att) using EfficientNetB3 and DenseNet201 as backbones.
• Integrated the Convolutional Block Attention Mechanism (CBAM) for refined feature map analysis.
• Implemented alternating backbone configurations to capture both high-resolution details and contextual features.

Main Results:

• The CNN-DP-Att model achieved a precision of 96%, surpassing standalone EfficientNetB3, DenseNet201, and other state-of-the-art models.
• The dual-path architecture effectively combined high-resolution and contextual feature extraction.
• The CBAM attention mechanism significantly improved feature refinement.

Conclusions:

• The CNN-DP-Att model demonstrates superior performance in classifying RBC morphology subtypes.
• This deep learning approach offers a robust tool for identifying erythrocyte abnormalities, overcoming manual microscopy limitations.
• The system holds potential for clinical laboratories in diagnosing and monitoring hematological conditions.