Atherosclerotic plaque segmentation using modified UNet with hybrid pooling layers

Atherosclerotic plaque segmentation is a vital task in cardiovascular image processing. Fuzzy nature of the carotid images makes it difficult to extract vital features from the plaque tissue region. UNet deep learning models use max-pooling layers for extraction of feature maps and are quite effective in medical image segmentation. In this study, we hypothesised that the UNet model with a hybrid pooling layer consisting of average pooling layer and max-pooling layers could exert more control on feature selection, and therefore be more effective solution for carotid plaque segmentation. We used a public database of 66 B-mode ultrasound images of the carotid artery for our experiments. We experimented with four cases of modified UNet model using a hybrid pooling layer with four different values of 'alpha' and compared it with the standard UNet model. Modified UNet model with hybrid pooling layers shows nearly 5% improvements in DSC and JI values.