TernausNet-based segmentation of cerebral infarction in magnetic resonance images

Objective: This paper aims to segment lesions effectively from magnetic resonance images (MRI) of cerebral infarction patients, using the state-of-the-art method of TernausNet. Methods: The subjects include 60 cerebral infarction patients diagnosed between May 2018 and September 2022. The MRI on their brains were collected, and the lesion areas were labelled by a team of experts. The use of multiple doctors in the first round of lesion labeling helps minimize bias, while any inconsistencies are addressed and corrected in the second round by a more experienced doctor. Next, the subjects were divided into a training set, and a test set by random, and the TernausNet, a novel deep architecture for segmentation, was called to segment the cerebral infarction areas from the magnetic resonance images. The segmentation results of TernausNet were contrasted with the expert labelled areas, with the aid of the difference plot. The performance of the network was measured by metrics like the proportion of specific agreement (PSA), volume similarity index (VSI) and Pompeiu-Hausdorff distance (PHD). Results: The segmentation results of TernausNet were in line with expert labelled areas, as 2.6%-10.8% of the data fell in the 95% limits of agreement. On the test set, the mean PSA was 0.67, the mean VSI was 0.76, and the mean PHD was 38.27 mm, all of which belong to the 95% confidence interval. Conclusions: TernausNet provides an effective tool to segment lesions from magnetic resonance images of cerebral infarction patients, shedding new lights on the clinical application of deep learning methods in brain image segmentation.