Our study has been published in the Frontiers in Digital Health.
By using a well-known pre-trained CNN model as a feature extractor, in combination with class activation maps, our aim was to understand which are the plaque areas, in carotid ultrasound images, that the model perceives in order to decide on a symptomatic or an asymptomatic plaque. The images mainly included the plaque, with some of the surrounding background present.
A difficulty arises when examples of uniformly echolucent (very dark) carotid plaques are given to the model.
There is room for further research and validation; very dark plaques in deep learning-based processes are often underrepresented.
You can read the study here.
