Tomek links-based SMOTE method for class imbalance in blood cell classification with dual path sliding window attention model.

Medical image processing has transformed the Complete Blood Count (CBC) analysis to enhance diagnostic accuracy and to detect diseases such as neurodegenerative diseases, infections, and anemia through non-invasive blood cell analysis. Nonetheless, the Multispectral Imaging (MSI) datasets have class imbalance, which severely impairs the existing deep learning models. To overcome this, we introduce a new Dual Path Sliding Window Attention (DP-SWA) model with a convolutional block to extract the initial feature, dual-path processing with local and global sliding window attention, and a fusion compression block. This architectural design is complemented with a hybrid SMOTE-Tomek Links data balancing strategy to eliminate the class imbalance. Our proposed model with 0.937 million parameters and 1.05 GFLOPs is able to achieve an inference time of 4.3 ms and a state-of-the-art accuracy of 99.02 % in the multi-wavelength MSI dataset, which is better than the best-performing baseline, YOLOv8-N (93.01 %), and other current models. DP-SWA is an effective method that addresses the issue of class imbalance and provides better accuracy with an impressive level of computational efficiency, which has great potential to increase the accuracy of clinical diagnosis.