Attention-Based Dual-Path Deep Learning for Blood Cell Image Classification Using ConvNeXt and Swin Transformer

In the rapidly evolving field of medical image analysis, the precise classification of blood cells plays a crucial role in diagnosing and monitoring numerous hematological disorders. Traditional methods, while effective, often require significant manual effort and expert knowledge, leading to potential delays and inconsistencies in diagnosis. Addressing these challenges, this paper introduces a groundbreaking dual-path deep learning architecture that synergistically combines ConvNeXt and Swin Transformer networks. This innovative approach leverages the strengths of convolutional neural networks for local feature extraction and transformers for global context integration, effectively capturing the complex morphological variations in blood cells. Furthermore, the incorporation of a Multi-scale Preprocessing Module (MPM) significantly enhances the image quality, employing techniques such as local contrast enhancement, global illumination normalization, and morphological feature enhancement to improve the visibility and differentiation of cellular features. Tested on a comprehensive dataset of 17,092 blood cell images, our model achieves an unprecedented accuracy of 99.98%, demonstrating superior performance over existing methods. This level of accuracy not only underscores the effectiveness of our model but also highlights its potential to serve as a reliable tool in clinical settings, facilitating faster and more accurate blood cell analysis. By automating the classification process with high precision, our approach promises to enhance diagnostic workflows, reduce the workload on medical professionals, and ultimately contribute to better patient outcomes in the field of hematology.