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

Accurate classification of red blood cell (RBC) morphology subtypes is crucial for diagnosing and monitoring various hematological disorders. However, the subtle and complex differences between cell types make this task challenging, and human interpretation errors are often inevitable. Ongoing research continues to advance this field, particularly through medical imaging and deep learning. This paper presents a novel CNN-DP-Att (dual-path convolutional neural network with CBAM Attention) architecture, exploiting EfficientNetB3 and DenseNet201 as the backbone networks, combined with the convolutional block attention mechanism (CBAM) for improved extraction of subtle features. The architecture alternates between two backbone configurations while maintaining identical attention and classification components. The first configuration captures fine-grained details with fewer parameters using EfficientNetB3, while the second extracts more complex features through the dense connectivity provided by DenseNet201. The term dual-path refers to the integration of high-resolution and contextual feature extraction paths within the same framework. While the first focuses on high-resolution spatial details, the other captures contextual representations. Both paths are merged for final classification. The CBAM attention mechanism refines feature maps by applying both channel and spatial attention. The CNN-DP-Att achieves the highest precision of 96% when compared against standalone EfficientNetB3 and DenseNet201 and other state-of-the-art models. This is due to the effectiveness of the dual backbone architecture when combined with the improved feature refinement attributed to the attention mechanism. This system accurately identifies and classifies numerous erythrocyte abnormalities, overcoming the limitations of manual microscopy and offering a useful tool for clinical laboratories.