Purpose
To develop and evaluate an artificial intelligence based semantic segmentation technique for an improved efficiency of the cumbersome post-processing in late enhancement (LE) imaging.
Methods and Materials
A convolutional neural network for semantic segmentation of endo- and epicardial borders was pre-trained using 100.248 cardiac cine MR images from the public Kaggle second annual data-science-bowl-database [1], and labels obtained using the method presented in [2]. Transfer learning was then used to qualify the network for the segmentation of left-ventricular images with LE contrast in short axis. LE images and manual segmentation labels from 216 patients with history of acute myocardial infarction of our local database were used for this purpose (~2.200 images in...
Results
The application of the neural network resulted in meaningful segmentation results for all 26 patients. The obtained Dice score was 0.76 ± 0.06. A trend towards slightly lower Dice scores for basal and apical images was apparent with respect to mid-ventricular slices. Figure 1 and 2 exemplarily show a comparison of the manual and the AI based segmentation for a high (Figure 1) and a low (Figure 2) Dice score, respectively.
Conclusion
Artificial intelligence based segmentation promises a breakthrough improvement of the efficiency in LE diagnostics. For technical reasons we fixed the matrix size of all images to 192 x 192 in our study. This significantly accelerated the training process for the presented proof-of-principle, however, it also required a reduction of the resolution of the images (and labels) in almost all cases. A repetition of the training at full resolution might also improve the accuracy of the segmentation. Furthermore, the inclusion of more training data will undoubtedly...
References
[1] Kaggle. Data Science Bowl Cardiac Challenge Data. https://www.kaggle.com/c/second-annual-data-science-bowl/data.
[2] Bai et al., JCMR 20:65 (2018)