Efficient CNNs and Energy Efficient SRAM Design for ubiquitous medical devices

Abstract

Intermittent monitoring of urinary bladder volume aids management of common conditions such as post-operative urinary retention. Urinary retention is prevented by catheterization, an invasive procedure that greatly increases urinary tract infection. Ultrasound imaging has been used to estimate bladder volume as it is portable, non-ionizing, and low-cost. Despite this, ultrasound technology faces fundamental challenges limiting its usability for next generation wearable technologies. (1) Current systems require skilled manual scanning with attendant measurement variability. (2) Current systems are insufficiently energy-efficient to permit ubiquitous wearable device deployment. We propose to develop an energy efficient system capable of real-time bladder volume monitoring. This system will incorporate several key innovations, including (1) Convolutional Neural Network (CNN) based segmentation algorithms employed to generate spatiotemporally accurate bladder volume estimates and (2) energy efficient static random access memory (SRAM) with in-memory dot-product computation for low-power segmentation network implementation. The aim is to develop platform technology embodiments deployable across a wide range of health-monitoring wearable device applications requiring accurate, real-time and autonomous tissue monitoring. We have selected bladder volume as the initial target application for development of these technologies.