Computer Aided Diagnosis for Contrast-Enhanced Ultrasound Using a Small Hybrid Transformer Neural Network

The Transformer Neural Network (TNN) architecture has emerged as a principal candidate in various data processing applications. Unlike traditional convolutional and recurrent Deep Neural Network (DNN) models, Transformers leverage a self-attention mechanism, allowing them to capture complex dependencies and relationships within data more effectively. This capability has enabled Transformers to consistently achieve superior performance across a wide range of tasks. Consequently, TNNs have become the architecture of choice for state-of-the-art models in both research and industry. This research aims to enhance AI-based automated diagnosis in medicine by exploring the small Hybrid Transformer Neural Network (HTNN) architecture for characterizing focal liver lesions (FLL) in contrast-enhanced ultrasound (CEUS) imagery. Our study initially reviewed recent TNN architectures employed in image classification tasks, followed by the identification of a suitable TNN variant for the aforementioned topic. This aim is supported by the fact that TNNs generally perform better with large training datasets, which is lacking in most medical datasets, including the one used in this study. We compared our approach with other machine learning solutions presented in the literature and found it achieves comparable accuracy. Additionally, our approach classifies more FLL types than most previous CEUS Computer-Aided Diagnosis (CAD) systems.