Improving blood cell classification in imbalanced medical imaging using Mixup-augmented graph-attention encoder-decoder network

Automated blood cell classification is crucial for improving the accuracy of hematological diagnoses, but it faces the ongoing challenge of severe class imbalance in multispectral datasets, where the number of red blood cells (RBCs) far exceeds that of clinically significant platelets (PLTs) and white blood cells (WBCs). To address this, we present a novel Convolutional Neural Network with Graph Attention (CNN-GA) model with an encoder-decoder architecture that includes an Adaptive Context Transfer layer, a Graph Attention Bridge, and multi-scale feature fusion for spatial-relational reasoning. This model uses a mixup-based minority-class balancing method to produce realistic synthetic samples for underrepresented classes while maintaining cellular morphology. With only 0.852 million parameters and 1.02 GFLOPs, our proposed CNN-GA model attains a state-of-the-art accuracy of 98.64% on the multi-wavelength dataset with an inference time of 3.8 ms. It constantly surpasses existing models across individual wavelengths and shows greater accuracy and efficiency than modern architectures, such as Swin-S (89.42 percent), and YOLOv9-C (94.61 percent). This work offers an effective and efficient solution for class-imbalanced blood cell analysis, thereby improving the detection of rare cell types in clinical settings.