Low-power AI processor designed for lung-sound monitoring
According to DeepTech, a team led by Wang Chao published a paper in the IEEE Journal of Solid-State Circuits describing a dedicated AI processor for lung-sound monitoring. The chip combines a two-stage hybrid neural network with reconfigurable processing units, dual-mode systolic dataflow and a multi-branch hybrid-driven memory architecture. The reported measurements include a minimum energy consumption of 2.76 pJ per operation and 0.39 mJ for processing one frame of normal lung sounds. The model was trained using multi-site cardiopulmonary auscultation records collected at Tongji Hospital in early 2020 from 172 confirmed COVID-19 cases and 50 normal participants. The chip was fabricated using a 55-nanometer process and has a core area of 4.39 square millimeters.
The report says the processor’s low-power design addresses the energy demands of long-term lung-sound monitoring. Its two-stage strategy and hardware architecture may also provide a reference for other physiological-signal monitoring chips, including those for heart sounds, electrocardiograms and electroencephalograms; the report does not establish clinical deployment or long-term monitoring outcomes.
时间回到 2020 年初的武汉,一台数字听诊器被送进同济医院新冠肺炎重症监护室,完成了极其艰难的第一手数据的采集。华中科技大学王超教授告诉 DeepTech,这台通过委托海外华侨捐赠的设备,在那一刻成为了连接生命的听觉窗口。
在六年后的今天,当年采集的这批珍贵的肺音数据,结出了一颗硬核的芯片果实。近日,王超团队基于这些真实的临床数据,在集成电路顶级期刊《IEEE 固态电路学报》上发表了历史首款面向肺音监测的专用 AI 处理器芯片论文。从听见新冠到造出来芯片,代表着算法和硬件结合迈出了实质性的一步。
#Huazhong University of Science and Technology#Tongji Hospital#IEEE Journal of Solid-State Circuits