Advancing Towards a Lower-Power Wearable Ultrasound Sensor for Real-Time Bladder Volume Monitoring using CNN Optimization

Recent advancements in sensor technologies have enabled the development of wearable ultrasound devices for continuous bladder volume monitoring. However, challenges related to power consumption, form factor, cost, and technical limitations associated with depth of signal penetration remain to be fully addressed. To tackle these challenges, this paper introduces a novel 11-channel wearable ultrasound sensor that leverages an efficient innovative convolutional neural network (CNN) framework and FPGA-based deep learning processing unit (DPU) accelerator to triple pitch the proposed sensor from 11-channel to 32-channel in real-time. This framework infers 32-channel full-array plane-wave channel datasets from a periodically sparse 11-channel array operating with as few as one-third of the elements in the full aperture. This device leverages a lightweight Conv-LSTM (LConv-LSTM) CNN model optimized for deployment on a single core of a DPUCZDX8G with B3136 architecture and has a balance between a high degree of parallelism and power efficiency. This device paves the way for transformative advancements in real-time bladder monitoring and management.