Machine learning techniques applied to medical imaging automated diagnosis is one of the most popular trends ever to reach the health market. FPGA technology allows to deploy custom hardware accelerators that are able to efficiently execute the most demanding deep learning algorithms, being a technology that has started to become available from the main cloud computing providers, e.g. Amazon, Baidu, Huawei…

At GL Research, a privately owned laboratory with a extensive expertise in FPGA technology, we have been working in the last three years in designing neuromorphic processing architectures that are able to evolve its own internal architecture to optimally match the requirements of the specific neural network models being executed. This is possible because of the inherent programmable nature of FPGA devices, allowing to squeeze the most computing power from a constrained electrical power budget. As a proof of concept, we trained our solution with a chest X-Ray image database and their associated diagnosis compiled from the Open Image database hosted by the U.S. National Library of Medicine.

Despite the fact of working on a very reduced and low resolution image set, the accuracy for several pathologies diagnostics matched the levels of the specialist practitioners, what encouraged us to launch a spin-off project aimed to serve the medical B2B and B2G markets. This project has been recently awarded with the MedTech 2017 price to the most disruptive business inititiative in Health Sciences by the Government of Navarra (SPAIN).

Relying in an external cloud providing artificial intelligence based diagnosis services is not an option for scenarios in which a continuous communication cannot be ensured, e.g. battlefield or natural disaster scenarios. In these situations, you need to move the intelligence right to the edge by embedding the neural network engine in a compact and low powered portable device.

In collaboration with Xilinx, the leading provider of All Programmable FPGAs, SoCs, MPSoCs and 3DICs, we have been working in migrating our FPGA based solution for the cloud to a single and highly embedded System-on-Chip device. Specifically, Xilinx provided us with early access to a ZCU102 development kit powered by a Zynq UltraScale+™ MPSoC device with a quad-core ARM® Cortex-A53, dual-core Cortex-R5 real-time processors, a Mali-400 MP2 graphics processing unit, 2,520 DSP slices and 600K fully reconfigurable system logic cells based on Xilinx’s 16nm FinFET+ programmable logic fabric.

By using the Xilinx reVISION machine learning software stack, we have been able to compress our neural network models using 32 bit floating point arithmetic into a highly efficient fixed point processing architecture using as low as 8 bit weights and 16 bit data. Once the neural network model have been successfully quantized, this can be deployed into the All Programmable MPSoC without a noticeable accuracy loss in the pathology prediction.

In conclusion, these highly programmable System-on-Chip devices packing FPGA and general purpose computing resources in a single chip will enable a new generation of battery powered automated diagnosis solutions with a huge resemblance with the Start Trek’s tricorder device.


Author: Javier Garcia

Coauthor(s): Dan Isaacs <>Director Connected Systems at Xilinx</>

Status: Work In Progress