Design Optimization of Integer Lifting DWT Circuitry for Implantable Neuroprosthetics

Abstract

Neuroprosthetics can benefit greatly from area and power efficient signal processing circuitry suitable for implanting alongside miniature neural probes that interface to the nervous system. This work identifies an optimal VLSI architecture for computing a I-dimensional multilevel discrete wavelet transform for multiple electrode channels simultaneously. The architecture is based on the lifting-scheme for wavelet computation and integer fixed-point precision for real-time processing under constraints imposed by implantability requirements. Two different computational node designs have been explored and compared to identify an optimal approach that minimizes power and chip area for a given number of levels and channels. Results demonstrate that on-chip computation is feasible prior to data transmission.

Analysis of Lifting and B-Spline DWT Implementations for Implantable Neuroprosthetics

“…2. For the Symmlet 4 basis, the set of lifting filters can be described as in Thomson et al [11] and Mason et al [12] by…”

“…This single hardware block can be repeatedly used to perform all the computational steps in Eq. 6 sequentially [11,12]. Sequential reuse of the same hardware reduces the area required by the overall DWT block without impacting performance in this low bandwidth application [12].…”

“…Recall that the relatively long period between neurologically relevant data samples motivated the design of a sequentially cycled single-stage CC. Analysis of preliminary DWT implementations [12] indicates that there is sufficient bandwidth between samples to process many data channels sequentially with a single CC, permitting significant savings in chip area.…”

Comparison of Lifting and B-spline DWT Implementations for Implantable Neuroprosthetics

“…3(a). A single hardware block can be used to perform each computational step sequentially, reducing the area required by the overall DWT block without impacting performance in this low bandwidth application [6]. This approach requires five cycles to compute results for one input sample pair.…”

B-Spline vs. Lifting Architecture for Neuroprosthetic Devices

“…The advantage is that a finite even integer multiplication can be performed by shifting and adding. Further, polyphase decomposition [12] is performed on the distributed part, Q(z) and R(z) to increase the hardware utilization. The traditional Mallat' s algorithm for evaluating the wavelet transform of a given signal involves recursively convolving the signal through two decomposition filters H and L, and decimating the result to obtain the approximation and detail coefficients at every decomposition level [5].…”