- A new deep learning method detects possible signals of Atrial Fibrillation recurrence with 90% accuracy.
- To do this, it analyzes and learns from three-dimensional cardiac MRI images and generates person-specific landmark-based anatomy representation.
According to CDC report, around 6.1 million people in the U.S suffer from Atrial Fibrillation (AFib) – an irregular and often rapid heart rate that raises the risk of heart failure, stroke, and other complications.
High blood pressure and advancing age account for up to 22% of AFib cases. Usually, it increases the risk of stroke by 5 times and causes 20% of ischemic strokes that occur when the flow of blood to the brain is obstructed by fatty deposits known as plaque in the blood vessel lining.
Although AFib is not life-threatening, it’s a serious condition that often needs emergency treatment. In many cases, it reappears after the treatment. To detect possible signals of its recurrence, researchers at the University of Utah have designed a deep learning method that detects AFib with 90% accuracy.
How It Works?
Deep neural networks use three-dimensional cardiac MRI picture to produce a person-specific landmark-based anatomy representation, eliminating tedious tasks like manual preprocessing and segmentation.
To predict AFiB recurrence, it analyzes the shaped of a left atrium (one of the four chambers of the heart) and looks for irregularities. However, the network cannot be efficiently trained with limited samples. Therefore, researchers applied a data augmentation approach to produce more statistically feasible information, and thus train the network while decreasing the risk of overfitting.
The convolutional neural network is trained on hundreds of MRI pictures, using NVIDIA Tesla GPUs with TensorFlow deep learning framework. Then they performed data augmentation on 75% of the original dataset to improve the network accuracy.
Reference: arXiv:1810.00475 | University of Utah
More specifically, the structure of left atrium exhibits a clustering in the shape space due to the vast number of possible arrangements of pulmonary veins. To deal with this number, they modeled left atrium shapes as a multi-model Gaussian distribution in the Principal Component Analysis subspace, with three components providing the best Bayesian information criterion.
Standard shape modeling vs proposed method | Courtesy of researchers
In this experiment, a total of 207 samples were used, out of which 175 were used for data augmentation and the remaining ones were sets aside for network testing (as unobserved samples).
Since the proposed technique works by learning shape descriptors from pictures, it has been utilized for automatic segmentation of left atrium with promising results.
After comparing this technique with the existing state-of-the-art shape analysis workflow that demands regular human intervention and correspondence optimization, they found that outcomes were statistically comparable. The recurrence predicted by the deep neural networks is 90% accurate with an error of ±0.06%.