Multimodal Deep Learning Model for Predicting Thrombus Thickness and Elasticity Using Ultrasound Imaging

In this study, we propose a multimodal deep learning model to predict the thickness and elasticity of blood clots (thrombi) using ultrasound imaging. We utilized an open dataset, Mus-V: Multimodal Ultrasound Vascular Segmentation, to train and evaluate our model. The model processes three types of ultrasound data: 1D ultrasound signals, 2D frequency-domain images obtained through Fast Fourier Transform (FFT), and Doppler ultrasound data. Each data type is analyzed using separate convolutional neural networks (CNNs) to extract important features. These features are then combined in fully connected layers to predict the thickness and elasticity of the thrombus. Our results demonstrate that combining multiple data modalities significantly improves prediction accuracy compared to models using a single data type. This approach highlights the potential for improving the early diagnosis and treatment of blood clot-related diseases through advanced multimodal deep learning techniques.