Image is Everything

Authors: Mike Barber, President and CEO, MICT and Jie Xue, President & CEO, MR, GE Healthcare

Ask any teenager, photographer, or radiologist whether image quality matters and they will all respond with a resounding “YES!”

For teens and patients alike, images can tell the story of who they are at a specific moment in time – and the greater the image quality, the more detail they can recognize. For clinicians and radiologists, images are an important tool for diagnosing disease, identifying the best course of treatment, and determining whether therapy is successful.

Image quality matters in healthcare – it can make the difference between identifying a small lesion in its early stages or finding that it has grown over several months or years.

Now, with new tools like deep learning at our disposal, we usher in the next generation in image quality and reconstruction.

But before we look forward, let’s take a moment to look back at image quality in CT and MRI and reconstruction over the past several decades.

In 1972, CT image reconstruction began with filtered back projection (FBP), a technique that was used over three decades, but challenged clinicians with image noise and artifacts in low dose settings.

With the desire to constantly improve, GE Healthcare pioneered iterative reconstruction in 2008. Still a popular technique used today, iterative reconstruction greatly improves image quality but, in its most advanced state, presents images that some have described as plastic-like, blurry or over-smoothed. Not to mention the long reconstruction time.

For MRI, image reconstruction hasn’t changed too significantly since its inception in the 1980s where a mathematical (Fourier) transform is applied to convert the discrete MR signals into images. Recent advances in accelerated MR imaging acquire less data and use alternative reconstruction techniques, such as parallel-imaging or iterative compressed-sensing reconstruction.

With new technology at our disposal, our GE engineers set out on a mission to improve image reconstruction once again, but this time in a completely different way.  

The idea was to use GE Healthcare’s Edison platform to build an algorithm that could use Deep Learning to reconstruct images. This technique leverages a Deep Neural Network (DNN), in which each layer is a series of mathematical calculations.

Think of a neural network like your brain, constantly taking in data and converting it into intelligence.

The CT algorithm’s “brain” was trained with data from high dose FBP phantom images and high fidelity FBP clinical images, under strict supervision to adjust and optimize the learning, to reconstruct high quality images while maintaining natural noise texture. The MR algorithm* was similarly trained on tens of thousands of images with supervised learning to recognize patterns of noise and low resolution in order to reconstruct the ideal object image. In addition, an innovative MR image ringing suppression technology is embedded in the reconstruction process.

The output in CT, TrueFidelity CT Images, solves the issue of having to choose between image quality and dose. They provide natural looking image texture at low dose – even in thin-slice images. And it’s done in less time than before. This workflow is considered clinically routine with less than a minute to reconstruct a whole heart and less than 90 seconds for an abdomen pelvis.

AIR? Recon DL* is designed to generate TrueFidelity MR images to improve signal-to-noise and image sharpness, enabling shorter scan times.  The goal is for clinicians and technologists to no longer have to compromise between image quality and scan time—an age-old problem with MRI.  This technology was designed to provide TrueFidelity MR images that are reconstructed from the raw MR data and are thus able to achieve high fidelity resolution, compared to other deep-learning image filtering methods.

In April, the FDA granted 510(k) clearance to our Deep Learning Image Reconstruction algorithm for TrueFidelity CT images, marking the first clearance of a Deep Learning technology for CT image reconstruction. And we’re proud to showcase AIR Recon DL as 510(k) pending at this year’s RSNA.

We are proud to once again usher in the next generation of image reconstruction – after all, we at GE Healthcare agree, image is everything.

GE Healthcare’s Edison offering comprises applications and smart devices built using the Edison platform. The platform uses an extensive catalog of healthcare-specific developer services to enable both GE developers and select strategic partners to design, develop, manage, secure and distribute advanced applications, services and AI algorithms quickly. Edison integrates and assimilates data from multiple sources, applying analytics and AI to not only transform data, but provide actionable insights that can be deployed on medical devices, via the cloud or at the edge of the device.

Visit gehealthcare.com for more information.

*AIR Recon DL is 510(k) pending at the US FDA and not available for sale.