Quick Fire Session
SCMR 22nd Annual Scientific Sessions
Hypertrophic cardiomyopathy (HCM) is characterized by unexplained left ventricular hypertrophy, myofibrillar disarray and myocardial fibrosis. An on-going clinical study  aims to improve the clinical risk stratification using cardiovascular MR with cine, native and postcontrast T1 mapping, and late gadolinium enhancement (LGE). Quantitative biomarkers will be extracted from the images, including left ventricular (LV) and right ventricular (RV) mass, volumes, wall thickness, mean myocardium T1 values, and scar percentage. Accurate segmentation of the epi and endocardial of the LV and RV from cine, LV from T1 and LGE are required for extracting these biomarkers. Manual segmentation is time-consuming and suffers from inter-observer variability. Deep convolutional neural networks (DCNN) have shown good performance in LV segmentation on cine images of normal and coronary artery patients . However, HCM patients often have irregular myocardial shapes and smaller blood pools, so that the segmentation is more challenging. After segmentation, the automatic extraction of biomarkers such as wall thickness requires automatic detection of myocardium segments. In this study, we aim to develop an automatic biomarker quantification method from multi-contrast MRI of HCM using DCNN-based segmentation followed by biomarker extraction.
The dataset includes around 2200 patients with the completed protocol. Manual segmentations and biomarker measurements from experts were performed. In order to maximize the segmentation performance, separate networks were trained on each contrast. For cine, a cascaded network structure based on a 3D DCNN was used to segment the epicardium of the LV and RV first, and then the masks were used to remove non-cardiac regions prior to segmenting the endocardium. The LV segments were determined by the intersection of epicardial contours of LV and RV with adjustments for even division. For T1 and LGE, segmentation was performed using individual 2D DCNN to segment epi and endocardial LV.
Table 1 shows the segmentation results of LV and RV at end-diastole and end-systole phases from cine images measured by Dice scores and the average perpendicular distances (APD). The Dice scores of epi and endocardial contours were 0.91 and 0.84 on native T1 images and were 0.94 and 0.92 on LGE images, respectively. Table 2 shows the symmetric mean absolute percentage error (sMAPE) in the corresponding biomarker quantifications. Figure 1 shows the automatic LV segmentation on LGE images.
A DCNN method was developed for automatic segmentation from multi-contrast MRI of HCM with automatic extraction of the related biomarkers. It shows promise in significantly reducing human effort and improving quantification accuracy, leading to reduced cost in clinical studies and improved analysis. Further studies will be performed to improve the segmentation and quantification accuracy and use advanced machine learning methods for analysis on the biomarkers.