Analog Convolutional Layer applied to Therapy-Driven Vagus-Nerve Detection

Abstract

Focused ultrasound vagus nerve stimulation (FUVNS) promises neurostimulation without implants, yet practical deployment is constrained by the need for accurate, real-time nerve localization and the cost, power, and operator burden of conventional imaging pipelines. This thesis proposes a complete hardware-software co-design for direct, on-chip localization tailored to FUVNS.

On the algorithmic side, we construct a task-specific ultrasound RF corpus spanning simulation phantoms and in-vivo acquisitions, and train a lightweight convolutional regressor that estimates nerve position directly from raw RF streams, bypassing B-mode reconstruction. Inter-channel correlation analysis guides an aggressive receive-element reduction, while hardware-aware quantization and magnitude pruning align model precision and sparsity with analog compute constraints, preserving inference robustness under stringent I/O and memory budgets.

On the hardware side, we design and simulate an analog convolutional layer based on charge-domain MAC operations, accepting analog activations from the ultrasound front-end and executing energy-efficient multiply–accumulate with compact passive structures. System- and block-level evaluations validate functional compatibility with the network’s dataflow and demonstrate substantial reductions in downstream bandwidth and front-end power compared with imaging-centric workflows.

Together, these contributions deliver a proof-of-concept pathway from RF acquisition to neurostimulation control within a single, application-specific edge-AI stack. The results indicate that direct-RF inference coupled with analog in-memory computation can enable compact, low-power vagus nerve localization systems.