|
|
Bicuspid Aortic Valve Malformation Classification with Weak Supervision
|
|
|
I worked on the project of detecting the bicuspid aortic valve malformation at Stanford.
Normal aortic valve has three cuspids, and due to congenital conditions and unhealthy lifestyles, it could
lead to a malformation of bicuspid valve.
And if left undetected and untreated, it could lead to cardiac arrest or worse.
The main difficulty in this project is that the condition has a prevalence of less than 2% in general
population, and only about 6% in our dataset.
On top of this, we are working with unlabeled datasets from the UK Biobank.
We applied domain knowledge on processed segments to form noisy
weak labels which is in turn used to train a deep neural network model (CNN-LSTM) to generate the final
predictions. In the final classification results, we have a 61% improvement in F1 score (37.8 to 60.8),
171% improvement in precision (30.7 to 83.3), and model improves as we add more unlabeled data!
|
|
|
|
|
|
Mitral Regurgitation Classification with Cardiac MRI Sequences
|
|
|
Mitral valve is the valve connecting the left atrium and left ventricle of the heart. Mitral valve
regurgitation is the condition where the blood flows back from the left ventricle into the left atrium.
And if left undetected and untreated, it will lead to severe complications including pulmonary edema,
blood clots, stroke, heart failure.
In the Mitral Valve Regurgitation classification pipeline, we employed a U-Net segmentation model and
applied the masks to the raw MRI datasets, then feed the center-cropped, masked dataset into a
discriminative deep neural network to get final probabilistic labels.
|
|
|
|
