Chip maker NXP Semiconductor is developing a neural networking technology that can identify features in chest Xrays to diagnose Covid-19.
The xAI neural network adds clearly defined errors into Xrays to create uncertainty that improves the accuracy of the network training. This is currently being tested on several neural network frameworks, including Google’s TensorFlow.
“With the unprecedented global Covid-19 pandemic situation, our xAI research teams believe that xAI might help enable the rapid detection of the disease in patients. It is still early days, but we are encouraged by the proof points we have seen,” said Lars Regers, CTO of NXP. “The use of CT radiology and X-ray imaging provides a fast alternative detection capability alongside the prescribed PCR testing and diagnosis protocols. CT and X-ray images could be processed by a suitably trained xAI model to differentiate between clean and infected cases. xAI allows for real-time inference confidence and explainable insights to aid clinical staff in determining the next stage of treatment.”
“Today we have multiple techniques for introducing the uncertainty, but we fit into TensorFlow as it fits into the bigger environment,” said Gowri Chindalore, Head of Technology & Business Strategy for Embedded Processors. The technique is helpful for improving the accuracy of machine learning (ML) algorithms not just for image recognition for Xrays but also for automotive ADAS and autonomous driving.
“This is at an early stage of development we are applying all three techniques to the data we have, such as traffic signs as well as Xrays, and we are looking at what lends itself best for model optimisation, ” said Chindalore. “Not every application really needs this uncertainty estimation, the majority of ML applications do well with an inference engine.”
Next: Embedding xAI in a microcontroller
“This is a pure software implementation at this point, using regular training and during the process we introduce the measurement of the uncertainly itself. We are introducing the uncertainty on top of the flow. Today these are computationally intensive and this is not an optimised inference engine and therefore the computation is run on a GPU or laptop,” he explained
In a hospital environment today the xAI model has to be run on a laptop next to an Xray machine.
“We are looking to open source some of the models we have built for Covid-19,” he said. “But we need data sets. We need hundred and hundred of images for training. If there are privacy issues we can offer the model to them with technical consultancy but if someone is willing to share we would be happy to work with them. We have started talking to universities in Europe and north America about running their data set on the model. There is also a group at Stanford [in the US] looking to pull this together globally and we are engaged in conversations with them,” he added.
At the same time NXP is looking at how it can implement the technology in an embedded chip in the next 12 month.
“If you really want to take it to the edge we need to be able to take this and quantise it to INT8 (8bit integer) rather than FP32 (32bit double precision floating point) which allows it to run on an ML accelerator in an application processor or microcontroller,” said Chindalore.
This quantisation process is not simple for the xAI technique. “If we quantise the uncertainty data you lose a lot of accuracy,” he said.
There are a couple of product lines with embedded ML accelerators that could be used, from the Cortex-A application processors to the Cortex-M microcontrollers. “Our Cortex-A GHz MPUs have machine learning processing around 2.3TOPS and our goal is to get xAI to the point where it can benefit from this, and then further optimise it for a Cortex-M7 1GHz microcontroller,” he said. “We are going to be building future MCUs with [ARM’s] U55 ML accelerator in 2021 and these will have up to 2.5TOPS of processing in future products.”
“Once this reaches the stage where collaborators pick it up, we would move the xAI to a GPU and then to the accelerator in the next 12 months,” he said.
Hospital researchers looking to use the model can contact the NXP research team by email.