DCT Learning-Based Hardware Design for Neural Signal Acquisition Systems

IEEE Trans. Circuits Syst. I

Türe

Baldassarre

et al. 2018

Self Cite

Implantable systems are nowadays being used to interface the human brain with external devices, in order to understand and potentially treat neurological disorders. The most predominant design constraints are the system's area and power. In this paper, we implement and combine advanced compressive sampling algorithms to reduce the power requirements of wireless telemetry. Moreover, we apply variable compression, to dynamically modify the device performance, based on the actual signal need. This paper presents an area-efficient adaptive system for wireless implantable devices, which dynamically reduces the power requirements yielding compression rates from 8× to 64×, with a high reconstruction performance, as qualitatively demonstrated on a human data set. Two different versions of the encoder have been designed and tested, one with and the second without the adaptive compression, requiring an area of 230×235 μm and 200 × 190 μm, respectively, while consuming only 0.47 μW at 0.8 V. The system is powered by a 4-coil inductive link with measured power transmission efficiency of 36%, while the distance between the external and internal coils is 10 mm. Wireless data communication is established by an OOK modulated narrowband and an IR-UWB transmitter, while consuming 124.2 pJ/bit and 45.2 pJ/pulse, respectively.Index Terms-Implantable integrated circuit, area-efficient, low-power, compressive sensing, neural signals, learning-based digital signal processing, signal recovery.

Section: B Learning-based Compressive Subsamplingmentioning

confidence: 99%

Adaptive Learning-Based Compressive Sampling for Low-power Wireless Implants

IEEE Trans. Circuits Syst. I

Türe

Baldassarre

et al. 2018

Self Cite

“…Aprile et al [8] proposed a DCT Learning-based compressive subsampling method for neural signals. They compared their method with several recent randomized sampling approaches, i.e., Bernoulli [2], Structured Hadamard sampling [9], and Multi-Channel Sampling [5] developed for compression of neural signals.…”

Section: Literature Reviewmentioning

confidence: 99%

“…Their proposed method improved the reconstruction quality compared to the considered baseline approaches. The authors in [8] used linear decoder in their work. However, in this paper, by utilizing a nonlinear deep learning decoder, our method outperforms their method and obtains better reconstruction results.…”

Section: Literature Reviewmentioning

confidence: 99%

“…We then employ the obtained sampling mask to compress all the intervals of the test signals. The reconstruction is then performed using the deep network described in Section V. Furthermore, we also consider the linear decoder (3) with the learned sampling mask as proposed in [8], we refer to this method as DCT-LB Linear. In addition, we also consider adaptive compression method explained in Section IV.…”

Section: Comparison To the Baselinesmentioning

confidence: 99%

See 1 more Smart Citation

Real-Time DCT Learning-based Reconstruction of Neural Signals

Mahabadi

2018 26th European Signal Processing Conference (EUSIPCO)

Cevher

2018

Self Cite

Wearable and implantable body sensor network systems are one of the key technologies for continuous monitoring of patient's vital health status such as temperature and blood pressure, and brain activity. Such devices are critical for early detection of emergency conditions of people at risk and offer a wide range of medical facilities and services. Despite continuous advances in the field of wearable and implantable medical devices, it still faces major challenges such as energy-efficient and low-latency reconstruction of signals. This work presents a power-efficient real-time system for recovering neural signals. Such systems are of high interest for implantable medical devices, where reconstruction of neural signals needs to be done in realtime with low energy consumption. We combine a deep network and DCT-learning based compressive sensing framework to propose a novel and efficient compression-decompression system for neural signals. We compare our approach with state-of-the-art compressive sensing methods and show that it achieves superior reconstruction performance with significantly less computing time.

“…However, such implementation does not apply any compression mechanism, requiring an important power budget. The LBCS technique has been applied on circuit implementation with DCT-based transform [13]. Even though its implementation shows great signal reconstruction performances, the actual hardware implementation, which requires relatively larger area and power consumption with respect to its LBCS-Hadamard counterpart, makes it more suitable for di↵erent application, such as image processing.…”

Section: Learning-based Compressive Subsamplingmentioning

confidence: 99%

An area and power efficient on-the-fly LBCS transformation for implantable neuronal signal acquisition systems

Proceedings of the 15th ACM International Conference on Computing Frontiers

Wüthrich

Baldassarre³

et al. 2018

Self Cite

A power and area e cient hardware encoding system tailored for wireless implantable applications is presented. Constant medical monitoring allowed by implantable devices is the most relevant alternative to current bulky monitoring systems, which, in case of severe mental diseases, require heavy surgery and long term hospitalization periods. In this work, the circuit design and the signal processing algorithm dovetail in order to allow real-time neuronal signal monitoring. Two main features must be met on the circuit level to facilitate the acceptance of the implant from the human body: small area and low power consumption. The presented work proposes a new compression scheme based on the Learning-Based Compressive Subsampling approach, which allows an area reduction with respect to recent published works, while allowing high signal reconstruction quality within low power requirements. The proposed method implements on-the-fly compression coe cients generation, which does not require large static memories. This new fully digital architecture handles the data compression of each individual neuronal acquisition channel with an area of 200 ⇥ 190µm in 0.18 µm CMOS technology, and a power dissipation of only 1.15µW .