Efficient biomedical image segmentation on Edge TPUs

Andreas M Kist, Michael Döllinger

Keywords: Edge TPU, image segmentation, U-Net, endoscopy

TL;DR: We suggest an optimization and deployment strategy to perform image segmentation solely on Edge TPUs

Abstract: Biomedical semantic segmentation is typically performed on dedicated, costly hardware. In a recent study, we suggested an optimized, tiny-weight U-Net for an inexpensive hardware accelerator, the Google Edge TPU. Using an open biomedical dataset for high-speed laryngeal videoendoscopy, we exemplarily show that we can dramatically reduce the parameter space and computations while keeping a high segmentation quality. Using a custom upsampling routine, we fully deployed optimized architectures to the Edge TPU. Combining the optimized architecture and the Edge TPU, we gain a total speedup of >79$\times$ compared to our initial baseline while keeping a high accuracy. This combination allows to provide immediate results at the point of care, especially in constrained computational environments.

Paper Type: methodological development

Primary Subject Area: Segmentation

Secondary Subject Area: Application: Endoscopy

Paper Status: based on accepted/submitted journal paper

Source Code Url: Based on https://github.com/anki-xyz/bagls, published example code within the original manuscript https://ieeexplore.ieee.org/abstract/document/9151951

Data Set Url: https://zenodo.org/record/3377544

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.