Abnormality Detection in Carotid Ultrasounds with Convolutional Networks

Carotid artery stenosis is a common disease responsible for roughly a quarter of all strokes. In carotid stenosis, plaque deposits in the carotid artery narrow the vessel and reduce or block blood flow, increasing the risk of stroke. In practice, a patient is diagnosed with carotid stenosis using a combination of gray-scale, color Doppler and spectral Doppler ultrasounds. We present two deep learning methods for automating gray-scale carotid ultrasound screening. The first is an object localization model that crops-out extraneous graphical and textual information in gray-scale ultrasounds. Our object-localization model achieves an intersection-overunion score of 92.1%. The second is a convolutional neural network trained to detect signs of stenosis in gray-scale carotid ultrasounds. We present a robust analysis of current architectures and their failings in being able to reason about current datasets; this points to a clear need for more granularly annotated data. As the first model trained on this dataset, this outcome serves to further influence the efforts of our radiologist partners.