A multidimensional transformer-CNN network trained with incomplete ultrasonic radiofrequency data of blood for red...

Evaluation of red blood cell (RBC) aggregation is crucial for early detection of diseases such as ischemic cardiovascular disease, type II diabetes mellitus, deep vein thrombosis, and sickle cell disease. Ultrasound, a non-invasive and real-time technique, is widely used for monitoring RBC behavior. However, measurement inaccuracies caused by instrumentation and human error can introduce data anomalies, degrading the generalization capability of deep learning models. To address this issue, we propose a Multidimensional Transformer-CNN (MTCN) model trained on incomplete data. Specifically, 20% of the original ultrasonic data is randomly masked via a Mask-Head module, followed by a multidimensional Transformer encoder. A multi-dimensional adaptive fusion module aggregates features across various dimensions, which are then passed through a classification layer. Considering that ultrasonic RF signals contain negative-valued components, we employ a Gaussian Error Linear Unit (GELU) activation function to preserve this information while ensuring model efficiency. Experimental results on RBC aggregation dataset demonstrate that MTCN outperforms existing models by achieving an accuracy of 96.89% and an F1-score of 96.92%. These findings confirm the model's robustness and strong generalization capability, providing a promising approach for the accurate and non-invasive monitoring of RBC aggregation.