An automatic segmentation method of left myocardium based on SSD model and CNN

Abstract

Automatic segmentation of left myocardium plays an important role in the diagnosis of cardiomyopathies. However, automatic segmentation of left myocardium in CMR images is challenging, since it is difficult to obtain high accuracy without manual intervention. In this paper, we propose an automatic segmentation method of left myocardium in CMR images by combing a Single Shot MultiBox Detector (SSD) model with a Convolutional Neural Network (CNN). The SSD model is used to detect the heart ROI, which will be applied as the input of the following CNN structure for myocardium segmentation. Our proposed method is evaluated on CMR images from Kaggle's Data Science Bowl Cardiac Challenge Data, where 1140 patients' CMR images are provided. The ground truth of left myocardium segmentation for these images are manually segmented by experts. Experimental results demonstrate the effectiveness of our method on automatically segment the left myocardium with the average DSC value equaling to 90.23%.