A multi-feature alignment fusion neural network model for red blood cell aggregation classification using ultrasonic radiofrequency data of blood.

Evaluation of red blood cell (RBC) aggregation is crucial for the early prevention and accurate diagnosis of diseases such as ischemic cardiovascular disease, type II diabetes mellitus, and sickle cell disease. Ultrasound technology is widely used in medical applications due to its non-invasive, real-time nature. Ultrasonic radio frequency (RF) signals echoed from tissues contain both regular and diffusion information about scatterers. However, models designed for RF signals often fail to fully capture the information within these signals, potentially reducing the accuracy of RBC aggregation classification. To address this, we propose a novel Multi-Feature Alignment Fusion Neural Network (MFAFNN) to improve RBC aggregation classification performance. The MFAFNN utilizes the Vision Transformer (ViT) to extract regular features from RF signals and employs the Multilinear Sequential Attention Network (MSAN) to capture diffusion features. To mitigate misclassification caused by fused features, we introduce Multi-Information Adaptive Fusion (MAF). Additionally, a loss fusion algorithm is incorporated to ensure the regular and diffusion features are effectively aligned while preserving multi-scale information. For our dataset, 13 blood samples with varying degrees of RBC aggregation, configured with 5% plasma concentration intervals, were used. Experimental results demonstrate that the MFAFNN outperforms existing models, achieving an accuracy of 94.95% and an F1-score of 94.87% on the test set. The proposed model enhances the classification accuracy of RBC aggregation using ultrasonic RF signals, offering a promising tool for non-invasive monitoring of RBC aggregation. Our code is available in https://github.com/ZZZlab-collab/Multi-feature-Alignment-Fusion-Neural-Network.git.