Consistency evaluation of an automatic segmentation for quantification of intracerebral hemorrhage using convolution neural network

Abstract

Objective To establish an automatic segmentation algorithm using convolution neural network, and to validate the consistency between the algorithm and manual segmentation. Methods One hundred andforty⁃six CT scans of intracerebral hemorrhage (ICH) were included from Chinese Intracranial Hemorrhage Image Database (CICHID). They were randomly divided into training set (n=90), testing set (n=26) and validation set(n=30). All CT scans were manual segmentation. Training set and testing set were used for algorithm training. The validation set was measured by four methods including manual segmentation, algorithm segmentation, accurate Tada formula and traditional Tada formula. The consistency test was performed. Results Compared with the Tada formula methods, the percentage error ofalgorithm values was the smallest 15.54 (8.41,23.18) %, and algorithm agreement with the manual reference was the strongest (correlation coefficient 0.983). Bland⁃Altman analysis showed that 93.33% of the data was within the 95% limits of agreement (95%LoA), and 95%LoA was narrow (⁃ 6.46-5.97 ml). No significant differences were found in size and shape (P > 0.05, for all). Conclusions The algorithm using convolutional neural network has a certain application prospect, but it needs still more validation in large sample research. DOI：10.3969/j.issn.1672⁃6731.2020.07.005

Keywords

• cerebral hemorrhage
• artificial intelligence
• neural networks (computer)
• tomography, x⁃ray computed