Dilation and Erosion for Left Atrium Scar Segmentation

Segmenting the left atrium (LA) scar from a magnetic resonance imaging (MRI) dataset is a challenging task because of the low quality of the source image and the uncertain location and the small shape of the LA scar. Recent studies always focus on optimizing the structure of the model while ignoring optimizing the preprocessing method when segmenting the LA scar. Since LA scars always appear around the boundary of LAs which is easier to predict, boundary information can be used as auxiliary information to help scar prediction. In order to alleviate the inaccurate prediction effect that the LA boundary brings, we use the expansion method to increase useful image foreground information. At the same time, since we only need to draw on the information of the LA boundary, the prediction inside LA is interfering with information for us. Therefore, we apply the corrosion method to remove the useless image foreground information. In summary, in this research, we investigate the effectiveness of two popular computer vision techniques, i.e., dilation and erosion, on the segmented LA, to help predict left atrium scar. Results on the LAScar dataset demonstrate they are useful as preprocessing approaches in related tasks.