1354941 - Autonomous cardiac field map segmentation for B0 shimming pipeline using a dual-modality Deep-learning model

B0 inhomogeneity is a long-lasting issue for CMR in high field (3T and above) scanners. B0 shimming is the standard way to improve the B0 field. However, today’s standard cardiac shimming protocol requires the operator to manually select a squared shim volume, which often falsely includes regions with large B0 deviation (e.g., Liver and chest wall). The flawed shim field can induce imaging artifacts and compromise the reliability of CMR protocols. In addition, recent hardware development of high-order B0 shimming requires a more precise shim volume to avoid overfitting from falsely included off-resonance sources. [1] In this study, we developed a deep learning-based cardiac shimming model that can reliably contour the cardiac region for B0 shim without human interaction. The model reached a high segmentation accuracy in B0 field maps with different acquisition parameters and showed high reliability in generating precise cardiac contours for cardiac shimming.

Methods:

We developed an automatic contouring model of B0 field maps for autonomous cardiac shimming. Different from the conventional models, two channels(magnitude and phase) of the B0 maps were used to improve the robustness of the segmentation.  The model is based on the UNet model^[1], and the pipeline is depicted in Fig 1. Briefly, After data pre-processing, the 3D deep learning model was trained with 5-fold cross-validation. The training set was composed of 3D field maps from 40 healthy subjects. Field maps covering the whole thoracic region were used.  In the testing set, seven subjects were included. Two imaging parameter sets (1.5x.15x3.5 mm³/ 3.5x3.5x5.6 mm³) and TEs(2.1ms/3.5ms) were used to test the generalizability of the models. The performance of the models trained with single (magnitude) and dual (magnitude-phase) modalities were compared.

Results:

Representative contours of the models are presented in Fig. 2. Compared to the magnitude image (panel A), the combined phase and magnitude images (panel B) showed enhanced contrast between the heart and the surrounding structures. A noticeable contour difference at the heart-liver interface is seen between the models. The false negative region from the conventional model(arrow) shows the models’ ill performance in differentiating the heart from the liver using only the magnitude images. The dice scores from the two models are compared in Fig. 3. The dual-modality model performs better than the model trained only on magnitude images with higher mean dice scores and lower inter-subject standard deviation(Single vs. Dual= 0.83±0.079 vs. 0.89±0.025, P< 0.05). Notably, the significantly reduced SD from the dual model shows a significantly improved reliability from the added information from the phase images.

Conclusion:

The proposed method incorporating the phase map effectively improves the segmentation accuracy in cardiac B0 field maps. This shows a promising direction in multi-modality deep learning-based MRI segmentation.