[2024] On-Chip Analog-based Detection of Vagus Nerve using Raw Receiver Channel Data

Ultrasound (US) vagus nerve stimulation (VNS) is a neuromodulation procedure requiring precise control over the location and intensity of a relatively high-pressure focal spot.  Medical imaging methods such as magnetic resonance imaging (MRI) and computed tomography (CT) have been used to guide the focal spot [1]. However, these methods are expensive to employ, present a low portability, and increase patient discomfort and risk due to their exposition to ionizing radiation. The development of a portable, low-cost, high-precision US neuromodulation system enabling image-guided neuromodulation procedures is of great interest to the medical community to mitigate such risk and to migrate towards home-based healthcare solutions. The continuous scaling of complementary metal-oxide-semiconductor (CMOS) technology has enabled remarkable advances in this direction, with works allowing for the steering of the focal spot electronically and even for the integration of both US imaging and stimulation ASICs within the same chip [2][3].

Efficient VNS therapeutic procedures rely on precise sonication protocol execution, which is vulnerable to human-induced errors during control-loop operation. Misalignment with the vagus nerve during sonication results in ineffective therapy and can possibly lead to surrounding tissue damage. Interpretation challenges arise with reconstructed 2D US B-mode images, particularly when targeting the vagus nerve through the muscle wall of the human neck. Lowering the sonication central frequency for better signal penetration compromises spatial resolution, complicating interpretation and increasing error likelihood [4]. Automated systems utilizing ML algorithms enable real-time vagus nerve recognition in US B-mode images, reducing algorithmic time overhead by leveraging pre-trained models [4-6]. However, existing solutions, reliant on high-level algorithms running on CPU or GPU, suffer from low portability and rely on bulk, power-hungry apparatus [4-6]. Integrating feature extraction closer to the sensor could enhance cost-effectiveness, power efficiency, and accessibility, improving both patient and medical professional convenience [7]. On-chip B-mode image reconstruction using digital blocks is impractical due to significant die-area and power consumption constraints. Utilizing raw US signal data as the ML model input circumvents frame-rate limitations imposed by B-mode image reconstruction.

Implementation of the multiply and accumulate (MAC) units – the base processing unit for hardware matrix multiplication - in the mixed-signal domain facilitates highly integrated and energy-efficient ML algorithms on silicon [8-10]. The significant quantization of weights and input signals to resolutions as low as 4 bits have no significant impact on the accuracy or inference error on ML models like artificial neural networks (ANN) [8]. This allows for compact hardware implementations using charge or current-domain MAC units. Additionally, these implementations exhibit robustness to CMOS device mismatch and process/temperature variations, as calibration steps can be embedded into the MAC units' weight training process, minimizing on-chip area requirements.

This work aims to explore the use of current and/or charge-based MAC implementations of an on-chip ML model for automated detection of the vagus nerve from raw ultrasound receiver channel signal samples. By integrating this inference directly on-chip, no bulk centralized processing units or cloud-based models requiring internet access are required, truly removing all barriers to the portability and operational complexity of future miniaturized US image-guided neuromodulation applications.

[1] Fomenko, Anton et al. “Low-intensity ultrasound neuromodulation: An overview of mechanisms and emerging human applications.” Brain stimulation vol. 11,6 (2018): 1209-1217. doi:10.1016/j.brs.2018.08.013

[2] T. Costa, C. Shi, K. Tien, J. Elloian, F. A. Cardoso and K. L. Shepard, "An Integrated 2D Ultrasound Phased Array Transmitter in CMOS With Pixel Pitch-Matched Beamforming," in IEEE Transactions on Biomedical Circuits and Systems, vol. 15, no. 4, pp. 731-742, Aug. 2021, doi: 10.1109/TBCAS.2021.3096722.

[3] A. Javid, C. Zhao and M. Kiani, "A 16-Channel High-Voltage ASIC with Programmable Delay Lines for Image-Guided Ultrasound Neuromodulation," 2022 IEEE Biomedical Circuits and Systems Conference (BioCAS), Taipei, Taiwan, 2022, pp. 419-423, doi: 10.1109/BioCAS54905.2022.9948689.

[4] Pashaei, Vida et al. “Flexible Body-Conformal Ultrasound Patches for Image-Guided Neuromodulation.” IEEE transactions on biomedical circuits and systems vol. 14,2 (2020): 305-318. doi:10.1109/TBCAS.2019.2959439

[5] A. F. Al-Battal, I. R. Lerman and T. Q. Nguyen, "Object Detection and Tracking in Ultrasound Scans Using an Optical Flow and Semantic Segmentation Framework Based on Convolutional Neural Networks," ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Singapore, Singapore, 2022, pp. 1096-1100, doi: 10.1109/ICASSP43922.2022.9747608.

[6] Nair, Arun Asokan et al. “Deep Learning to Obtain Simultaneous Image and Segmentation Outputs From a Single Input of Raw Ultrasound Channel Data.” IEEE transactions on ultrasonics, ferroelectrics, and frequency control vol. 67,12 (2020): 2493-2509. doi:10.1109/TUFFC.2020.2993779

[7] J. Van Assche, M. F. Carlino, M. D. Alea, S. Massaioli and G. Gielen, "From sensor to inference: end-to-end chip design for wearable and implantable biomedical applications," 2023 IEEE Biomedical Circuits and Systems Conference (BioCAS), Toronto, ON, Canada, 2023, pp. 1-5, doi: 10.1109/BioCAS58349.2023.10388515.

[8] Liu, Shih-Chii et al. “Prospects for Analog Circuits in Deep Networks.” ArXiv abs/2106.12444 (2021): n. pag.

[9] Y. Chen, Z. Wang, A. Patil and A. Basu, "A 2.86-TOPS/W Current Mirror Cross-Bar-Based Machine-Learning and Physical Unclonable Function Engine For Internet-of-Things Applications," in IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 66, no. 6, pp. 2240-2252, June 2019, doi: 10.1109/TCSI.2018.2889779.

[10] M. Carminati, S. Di Giacomo, M. Ronchi, G. Borghi and C. Fiorini, "Bringing In-Sensor Intelligence in Radiation Detectors: a Short Review," IEEE EUROCON 2023 - 20th International Conference on Smart Technologies, Torino, Italy, 2023, pp. 142-146, doi: 10.1109/EUROCON56442.2023.10198991.

Assignment

1^st part:  Literature review of image guided focused ultrasound neuromodulation and multiply and accumulate circuits for ML.

2^nd part: Design and simulation of an on-chip multiply and accumulate-based ML processing unit for the automatic detection of the vagus nerve via ultrasound.

3rd part: If time allows, tape-out the circuit and perform electrical and acoustic testing.

Requirements

MSc students from Microelectronics with focus on analog and mixed-signal IC design.

Interested students should include their CV, the list of courses attended, and a motivation letter, and send it to Tiago Costa (t.m.l.dacosta@tudelft.nl) and Diogo Dias (d.a.dias@tudelft.nl).

Contact

Last modified: 2024-04-23