Modular convolutional neural networks for adaptable ultrasound sensing and echo analysis

Using convolutional neural networks (CNNs) to process ultrasound signals is a widely studied research field with applications in nondestructive evaluation, medical diagnostics, and remote sensing. While some studies have demonstrated the ability of CNNs to classify target objects by analyzing the ultrasound echoes they produced, past approaches rely on fixed, monolithic network architectures that cannot be easily expanded to learn new objects after the initial training. In this presentation, we introduce a modular CNN approach for adaptable ultrasound sensing. Our approach uses a set of small CNNs, each specialized in distinguishing a particular target object by analyzing the echoes the object produced. The received echoes are inputs to specialized CNNs, where each CNN outputs a classification probability, i.e., the likelihood of the shape of the object that produced the input echoes. The object with the highest probability is used as the final prediction. We also demonstrate the adaptability of this modular architecture, which allows expanding the pool of known objects by simply adding CNNs specialized in identifying additional objects. This adaptability shows remarkable perception accuracy despite its simplicity. Additionally, specialized CNNs are further analyzed to understand distinct echo features that determine the object identification.