A Hybrid Memristor and Tunneling FET Based Platform for Low Power Medical Imaging Acceleration

Active Sites

Duke University; University of Notre Dame

Motivation

With the advancement in medical experimentation, imaging and sensing technologies, modern medical research and practice are inundated with structured and unstructured data at massive scales/sizes, and with such massive data, new issues emerge in modern medical research and practice. For example, even a two-dimensional (2D) image slice can contain millions of immune cells, and to trace and study various disease dynamics (e.g., cancer metastases), four-dimensional (4D) image data are frequently employed from many individual patients. Traditionally, pathologists use microscopy to examine cell images at the scale of tens or hundreds of cells. However, traditional hardware cannot possibly capture and reveal patterns and trends when millions or even billions of cells are involved, due to the lack of computational power. Thus, new computational platforms for modern medical research and practice are desperately needed. We propose to address this challenge through a novel deep learning platform built with new devices including tunneling FETs and memristors, which features ultra-low-power and high-performance.

Objectives

We will consider how to best design the platform to meet the computational needs in medical imaging.

Team

Yiyu Shi, Michael Niemier, and Sharon Hu (CSE, ND) will be responsible for the components using Tunneling FET; Yiran Chen and Hai Li (ECE, Duke) will be responsible for those with memristors; Danny Chen (CSE, ND) is responsible for the target algorithm and application.

Deliverables

The deliverables include preliminary simulations and comparisons of various possible architectures and configurations and the potential optimization, and benchmarking using medical imaging dataset on the impact in terms of the power, performance, and/or accuracy.

Experimental Plan and Industrial Relevance

We will compare our platform to existing solutions (CPUs, GPUs, FPGAs, etc) in terms of power, delay, speed, etc. assuming the same best-performing algorithm is implemented.  The project will be of significant interest to companies from healthcare sectors.

Milestones and Time-to-Completion

The estimated duration of this project is 1 year, with the following milestones: Q1: Evaluation and determination of the computational needs; Q2, Q3: Configuration of various architectures for comparison/optimization; Q4: Final power, performance and accuracy benchmarking.

Number of Graduate Students Supported

2

Budget

$100K/year

Total Cost to Completion

$100K