Synthesizing fMRI scans using EGG data to detect Alzheimer’s and other brain diseases

By far, (fMRI) Functional magnetic resonance imaging is the most popular method used in neurodiagnostic in contrast to such as PET, magne-toencephalography , preoperative electrical stimulations, functional impairment induced by Wada test, and lesions.

An imaging technique that uses magnetic fields to detect changes in cerebral blood flow as a marker for brain activity. Specifically, fMRI measures deoxygenated to oxygenated blood ratio in the brain (which has different magnetic susceptibility) to identify neurons that are firing (active neurons consume more oxygen), revealing which structures of the brain are active at a given moment in time.

A standard fMRI voxel of 55 mm3 (3–5 mm in each dimension) contains 5.5 mil-lion neurons, 2.2–5.5 × 1010 synapses, 22km of dendrites, and 220 km of axons.

Despite its widescale clinical adoption, this technology in the space of preventive healthcare is limited

Primary reasons being-

1. Affordability

standard fMRI costs more than 400,000€ and around 600€ per hour, this is extremely expensive but needed giving how fMRIs are useful to detect a huge variety of issues and diseases.

2. Accessibility

Developing countries, where neurogenerative diseases like Alzheimer are predicted to exponentially grow with need the urgent preventive medical care to prevent a “Alzheimer’s time bomb” as it is often referred to in the UK. And it is yet the start reality is that millions of patients slip into the clutches of these unfortunate diseases day in and out Developing countries like Nigeria cannot afford to have sufficient fMRI unit supply in comparison to demand. This limited number cannot scan every person with higher degree symptoms of neurodegenerative diseases let alone patients with early-stage symptoms giving way to brain health pandemic of millions.

3. Inconvenient

Lying still inside a large machine is not a pleasing experience. The large form factor is often a source of inconvenience for patients causing clostrophobia. Also, for high-resolution spatial imaging remaining still is a necessity. It is harder for those with clinical problems to stay still for long.in awake subjects, subtle movements are always identified across images along time. The motion amplitude and its spatial coherence and synchrony with the paradigm may be responsible for significant signal changes. Using head restraints or bite bars may injure epileptics who have a seizure inside the scanner; bite bars may also discomfort those with dental prostheses.

Consequences of problem

There is growing evidence that diseases like Alzheimer can be prevented if detected early and yet the current healthcare industry remains reactive. The inefficiencies of the widely accepted diagnostic method further concretize the gap in quality healthcare for preventable conditions.

We at Synthesia, asked: how can we innovate creatively to deliver accurate quality preventive healthcare at a marginal cost to people who need it most?WHAT if we could deliver the accuracy of fMRI through a simple EEG electrode cap?!

We borrowed from the booming AI industry to help us answer these questions- questions which if answered correctly and with care of attention could help millions of people across the globe live a life of authenticity devoid of any burdens on their identity padded by memory loss, seizures, etc. Our solution could allow the detection of every person that would be at risk with an 800€ EEG cap instead of a 400,000€ fMRI scan.

For the innovation part, we marveled at the amazing state-of-the-art use cases of WGAN algorithms and transferred their applications to medical imagery. While looking at the latest data releases we noticed a growing trend of recording EEG + fMRI, the scientific community is trying to use EEG data to improve fMRI accuracy. But we asked isn’t the other way around needed? So we just a crazy what-if idea: “What if using a GAN we could reconstruct a fMRI scan using only EEG data”

This was crazy

But after some long whiteboard and discussions with researchers, we developed more and more the idea. This precise crazy idea, this moonshot gave Synthesia.

How our technology is working?

How do we do it? how can we generate artificial fMRI scans with only simple EGG data?

Indeed, this is a challenging task given the different methods as well as the resolution type of images generated by EEG and fMRI. Unlike EEG, f MRI does not directly measure neural activity; instead, it relies on changes in oxygenation, blood volume, and flow. Further, Because of the hemodynamic lag — the amount of time it takes for local blood-oxygen levels to rise and peak — the temporal resolution of f MRI is limited to several (1–6) seconds. As a result, EEG has a very high temporal resolution and low spatial resolution, and fMRI has a high spatial resolution but low temporal resolution.

To combine the two aspects, we used a GAN. It is a set of two artificial neural networks competing in a zero-sum game to win. We more specifically used a WGAN, to improve the quality and accuracy

Data preparation

We’re first doing some little modification to our data to allow our WGAN to work well, we’re performing an STFT (Short-time Fourier transform) on the EEG data and downsampling a bit the MRI data.

As our data doesn’t have the same temporal resolution, we need to “sync” them with a process that we’re calling BOLD shift emulation. We have a full 3D snapshot of the brain every 5 seconds, we are matching 5 seconds of EEG data with 1 fMRI snapshot

This is the diagram of how our WGAN is working:

Our GAN is composed of two main components: a discriminator and a generator.

The goal of the Discriminator is to recognize instances synthesized by the Generator, if it is able to do so then a penalization is given to the Generator. On the other hand, if the Discriminator does not recognize those synthesized instances, a penalization is given to Discriminatorn itself.

Here we are first passing EEG data and fMRI through a Dense and a GRU (a specific type of ANN) and we’re using the l_c learning function to map the temporal relationship between the data.

We’re after passing our data through a decoder and comparing it with the ground truth fMRI data and computing the L-r loss function.

With this process, we’re able to have an artificially synthesized fMRI scan from our EGG data

Once our fMRI is synthesized we’re using a CNN to detect a low neural activity in the hippocampus. to train it we’re passing the scam through 2 convolution and pool layers and passing it through a mish activation function. To backpropagate the weight correction on the CNN we’re comparing our prediction with the labeled data and using it as a softmax function.

Our data

To get clinical level accuracy we need to redefine the standards of fMRI + EEG data. That’s why we will collect data on the brain of 300 patients with special fMRI compatible electrodes. On this data, we will perform denoising on the EEG.

Our headset:

To capture high quality EEG recording within a 20minutes time frame. For this, we’re using commercially available dry electrodes.

Those are able to capture low noise data without requiring gel or any specific setup.

How we commercialize it:

We are first selling our product in two markets. The UK and Nigeria. It will allow synthesis to tackle two different problems. The no early diagnosis in the UK and no fMRI problem in Nigeria.

Our subscription plan will cover the headset and cloud detection. The price will be 300$ per month in the UK and 3$ in Nigeria.

We at Synthesia are confident that our technology, values and mission will have a huge positive impact on the world and that our what if idea will improve people’s life.