Object detection inside a wearable ultrasound neuromodulator patch

Abstract

Neuromodulation of the vagus nerve is used as a treatment for all kinds of ailments and even as a means of improving the wearer's physiology, however, this form of treatment is not popular due to its invasive nature, high chance of side effects, and short period between reimplementation surgery, as such an alternative is sought in the form of neuromodulation using ultrasound. In this thesis, 2 designs for wearable ultrasound neuromodulators for the vagus nerve are suggested, based on these designs a solution is made, solving the problem of detecting vagus nerve within a wearable environment. To detect the vagus nerve two methods are proposed: neural networks and template matching. Based on these methods and these proposed designs 5 unique works have been created, Vivo las vagus (VLV) an object detector using a neural network, a mobile implementation of both VLV as well as template matching, an FPGA implementation of template matching, and 2 FPGA implementations of VLV in which one uses a streaming dataflow architecture and the other a systolic array architecture.

The best results are achieved with the streaming dataflow architecture implementation of VLV within an FPGA, resulting in an accuracy of 87.5 percent on the test set with the 10.88 FPS/watt, and inference of 0.174 seconds. This was achieved by using FINN, a community project for converting software neural networks into HDL representation for the FPGA. Combined with to the best of the author's knowledge, first-ever created loss function to automatically decrease the bit width of a quantized neural network layer without impacting the accuracy during training creating the first-ever fully automated end to end flow for creating a software neural network object detector and converting it towards an HDL representation, allowing biomedical engineers without knowledge of digital electronics or Neural networks to simply load in data and run the python files. To assess the accuracy an accuracy calculation function was created together with a dataset and test set with images taken from [1]. As the dataset has shown to be lacking severely in variety the accuracy assessment of all of the implementations can be considered moot. The VLV FINN implementation was compared to other FPGA implementations based on energy efficiency, showing that the work created within this thesis is one of the best in terms of power efficiency and the smallest in terms of resource usage footprint.