Enabling Ultra-Low Power Ultrasound Imaging with Compute-in-Memory Sparse Reconstruction Accelerator
- 19 October 2023
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 2, 1-5
Abstract
Medical ultrasound has played an integral role in the diagnosis of diseases over the past decades. Low power consumption of portable ultrasound is critical for its application in remote and point-of-care clinical scenarios. Among power reduction techniques, subsampling the ultrasound analog frontend channels is effective for reducing power, and sparse reconstruction is needed to guarantee final image quality. Although many successful reconstruction algorithms exist, including deep learning methods, we observe that locally executing these workloads by conventional hardware platforms cannot bring power savings for portable ultrasound. In this work, we aim to enable ultra-low power ultrasound imaging by approaching from both the algorithm and hardware fronts. For sparse reconstruction, we design a compact neural network that allows at most 8× channel subsampling while only using 11.7% operations of the baseline. We also design a digital compute-in-memory (DCIM) accelerator to process the model in portable ultrasound. The proposed row-centric inter-layer pipelining addresses the challenge of accelerating U-Net models, reducing the amount of required buffers by 4.6×. With an energy efficiency of 27.8 TOPS/W at 28 nm, the DCIM accelerator provides up to 57.8% power savings for portable ultrasound.Keywords
Funding Information
- Georgia Institute of Technology
This publication has 38 references indexed in Scilit:
- Deep Convolutional Neural Network for Inverse Problems in ImagingIEEE Transactions on Image Processing, 2017
- Highly functional and reliable 8Mb STT-MRAM embedded in 28nm logic2016
- Eyeriss: An Energy-Efficient Reconfigurable Accelerator for Deep Convolutional Neural NetworksIEEE Journal of Solid-State Circuits, 2016
- Scaling mixed-signal neuromorphic processors to 28 nm FD-SOI technologies2016
- Fused-layer CNN accelerators2016
- Deep Residual Learning for Image Recognition2016
- U-Net: Convolutional Networks for Biomedical Image Segmentation2015
- A 16-channel 38.6 mW/ch fully integrated Analog Front-End for handheld Ultrasound imaging2014
- Review article: Use of ultrasound in the developing worldInternational Journal of Emergency Medicine, 2011
- Image quality assessment: from error visibility to structural similarityIEEE Transactions on Image Processing, 2004