Background and objective: Segmentation of the left ventricular (LV) myocardium (Myo) and RV endocardium on cine cardiac magnetic resonance (CMR) images represents an essential step for cardiac-function evaluation and diagnosis. In order to have a common reference for comparing segmentation algorithms, several CMR image datasets were made available, but in general they do not include the most apical and basal slices, and/or gold standard tracing is limited to only one of the two ventricles, thus not fully corresponding to real clinical practice. Our aim was to develop a deep learning (DL) approach for automated segmentation of both RV and LV chambers from short-axis (SAX) CMR images, reporting separately the performance for basal slices, together with the applied criterion of choice.
Method: A retrospectively selected database (DB1) of 210 cine sequences (3 pathology groups) was considered: images (GE, 1.5 T) were acquired at Centro Cardiologico Monzino (Milan, Italy), and end-diastolic (ED) and end-systolic frames (ES) were manually segmented (gold standard, GS). Automatic ED and ES RV and LV segmentation were performed with a U-Net inspired architecture, where skip connections were redesigned introducing dense blocks to alleviate the semantic gap between the U-Net encoder and decoder. The proposed architecture was trained including: A) the basal slices where the Myo surrounded the LV for at least the 50% and all the other slice; B) all the slices where the Myo completely surrounded the LV. To evaluate the clinical relevance of the proposed architecture in a practical use case scenario, a graphical user interface was developed to allow clinicians to revise, and correct when needed, the automatic segmentation. Additionally, to assess generalizability, analysis of CMR images obtained in 12 healthy volunteers (DB2) with different equipment (Siemens, 3T) and settings was performed.
Results: The proposed architecture outperformed the original U-Net. Comparing the performance on DB1 between the two criteria, no significant differences were measured when considering all slices together, but were present when only basal slices were examined. Automatic and manually-adjusted segmentation performed similarly compared to the GS (bias±95%LoA): LVEDV -1±12 ml, LVESV -1±14 ml, RVEDV 6±12 ml, RVESV 6±14 ml, ED LV mass 6±26 g, ES LV mass 5±26 g). Also, generalizability showed very similar performance, with Dice scores of 0.944 (LV), 0.908 (RV) and 0.852 (Myo) on DB1, and 0.940 (LV), 0.880 (RV), and 0.856 (Myo) on DB2.
Conclusions: Our results support the potential of DL methods for accurate LV and RV contours segmentation and the advantages of dense skip connections in alleviating the semantic gap generated when high level features are concatenated with lower level feature. The evaluation on our dataset, considering separately the performance on basal and apical slices, reveals the potential of DL approaches for fast, accurate and reliable automated cardiac segmentation in a real clinical setting.
Keywords: Cardiac segmentation: Cine cardiac magnetic resonance; Convolutional neural networks; Dense skip connections.
Copyright © 2021 Elsevier B.V. All rights reserved.