Materials and Methods

A total of 44 subjects (29 males and 15 females; age: 23.5 ±3.4 yr; TOF: 25, control: 19) participated in the study approved by the institution’s review board after providing informed consents. The study was performed on a 3.0T MRI system (Skyra, Siemens) in Kaohsiung Veterans General Hospital. We acquired modified Look-Locker inversion recovery (MOLLI) T1 images using the parameters (slices: mid-cavity and/or apical and/or basal, MOLLI acquisition scheme: 5(3)3, TR/TE: 295.26/1.12ms, matrix: 256×152). Totally 123 MOLLI data sets were analyzed in this study. We implemented SegNet³ (Figure 1) using the TensorFlow framework and the Python environment. The SegNet is a convolutional neural network consisted of 26 convolutional layers. The output layer is a 4-class soft-max classifier. The four classes were left ventricle walls (LVw), blood pool region (LVb), and right ventricle (RV) and background. The input images were the T1 maps reconstructed from MOLLI data sets and the output images consisted of manually labeled LVw, LVb, and RV regions. The data sets of subjects were split into two subsets, subset 1 (control: 15 and TOF: 8 subjects, 60 images) and subset 2 (control: 14 and TOF: 8 subjects, 63 images). The two subsets were used to train the SegNet and cross-validate the performance. The optimization of SegNet were 50,000 back-propagation iterations with 6 training images evaluated per iteration. For comparison, the dataset underwent the segmentation method, layer-growing, presented in ISMRM 2017 ^1,2. In short, the method consisted of a procedure to obtain a synthesized image with minimum blood signal and a layer-growing algorithm to generate an LVw mask. We evaluated the performance of SegNet and layer-growing by a dice coefficient [dice= (2(X∩Y))/(X+Y), X: a reference mask and Y: a mask generated by SegNet or layer-growing]. We used the obtained RV and LVw masks to identify the anterior and inferior boundaries of the septal walls and distributed LVw masks into the AHA segments.

Results

Figure 2 displays the segmented results of the three slices of a subject. The obtained masks (red: LVb, green: LVw, blue: RV) overlaying on T1 maps were generated by layer-growing and SegNet (500, 5000, 50000 iteration steps), respectively. Notice that SegNet with more iteration steps prominently improve the accuracy of the segmentation results. For this subject, the dice coefficients of the labeled heart regions (LVw, LVb, RV) were (0.89, 0.96, 0.95) and (0.80, 0.93, 0.83) for the optimized SegNet and layer-growing methods, respectively. Figure 3 shows the obtained AHA segments of this subject. Table 1 lists the averaged dice coefficients. In comparison to layer-growing, the SegNet method significantly improved the segmentation accuracy in LVw and RV (p < 0.01, paired t-test). In addition, the dice coefficients were lower in apical slices processed by the SegNet method.

Discussions and Conclusion

In our previous study, we implemented a fully automatic bullseye analysis for native T1 values of the myocardium. Accurate segmentation results allow reliable analysis of regional T1 values. In this study, we used the recently emerging deep learning for the segmentation task. The results supported that SegNet outperformed the layer-growing method. The average dice coefficient of whole heart was approximately 0.97. The results support the trained SegNet could be a robust tool for whole-heart extraction. We aim to release this tool since it could be as useful as the brain extraction tool (BET) provided by FSL (FMRIB Software Library, the Analysis Group, FMRIB, Oxford, UK.). The major restriction of this study was that the training and validation data sets were limited to control and TOF subjects. Increasing the variety of the training data sets of SegNet could lead to more accurate segmentation and widen the clinical applications. In conclusion, the SegNet method increasing the accuracy of segmentation could improve the reliability and consistency of clinical applications of myocardial T1 values.

Acknowledgements

Supported by the Ministry of Science and Technology under grants 104-2314-B-011-001-MY3