Using Deep Learning to Fight Heart Disease

Using Deep Learning to Fight Heart Disease


The heart [archival audio] There operates within the human body a marvelous machine For something we talk about so
much, it’s remarkably hard to know what’s actually happening in your own heart So the idea is that in order to diagnose heart disease, you either have to do some
sort of invasive process like open-heart surgery or do some sort of
catheter to go up into your heart to look at the how the structures if
there’s any blockages or anything like that Another way is to do imaging, so you
can do MRI or CT imaging NARRATOR: The problem with imaging though is that radiologist workloads have been steadily increasing over the years. One of the time-consuming
tasks radiologists have to do is left ventricle segmentation MAI: When they look at
the images they will try to look at the left ventricle of the LV which is the
largest chamber in the heart and the health of the LV is a good indicator of
heart health in general But the task of looking at the LV—measuring the LV—requires manual or semi-automatic segmentation of the LV. So a cardiologist or
radiologist or both have to look at the image and trace the outline of the LV
and that’s a very tedious, time-consuming, error-prone process NARRATOR: To process the MRI
scans can take an expert from 30 to 60 minutes for a single patient. Why does it take so long? the images that you look at they can be very grainy and there can be
a lot of occlusions. Different parts of the heart will cover the LV
depending on where in the cycle you are and where in the heart you’re taking the
picture NARRATOR: Contouring requires deep knowledge of the heart’s anatomy which also means it can be challenging to train people to trace the LV manually And so what we want to do is apply a deep learning method to it. Specifically
we use a deep learning model called a U-Net so essentially given an image of a
heart, the model has to output a mask indicating where the LV is in that image NARRATOR: The beauty of deep learning is that there’s no need to teach the neural
network heart anatomy Once a neural network is trained on enough
professional contours it implicitly learns the features of the heart
relevant to segmentation Once you have the trained model, then you can
have it running 24 hours a day seven days a week. It doesn’t get tired, it
doesn’t get bored, so it can be very efficient once you have a trained model. NARRATOR: Manual contouring can take from 30 to 60 minutes Automated segmentation cuts that
down to less than a second Patients are better served by doctors who aren’t
overworked, fatigued, and prone to making errors Deep learning models like the one
Mai is developing can help doctors cope with increasing workloads, and improve
patient outcomes for the 28 million Americans diagnosed with heart disease