Introduction: Segmentation is an image processing technique that partitions an image into segments. There are many types of segmentation that differ in purpose, execution and method. Use of segmentation on traumatic brain injury patient images is a method to quantify the volume of intracranial hemorrhage, which is a crucial metric during treatment. Volume of intracranial hemorrhage is the most significant predictor of 30-day mortality. For an initial assessment, non-contrast head CT is the chosen method of imaging. Purpose: Provide a quantitative assessment of the volumetric change in traumatic brain injury patients with a intracranial hemorrhage; compare accuracy of different segmentation methods; determine which method is most time efficient; analyse the trend between quantitative volumetric change and assessed volumetric change on radiological reports. Methods: This study used the retrospective experimental method. From the imaging database at University Medical Centre Ljubljana, we selected five patients in whom traumatic intracranial hemorrhage had been confirmed by non-contrast head CT. We included a total of 48 CT scans in this study. All patient data had been anonymized beforehand. The study was conducted using open-source software 3D Slicer 5.8.1 with extensions in which we segmented the images manually, semi-automatically and automatically. Radiological reports were converted into volume estimates using a semiquantitative scale. To compare the spatial overlap between segmentations, we used the Dice coefficient and the Jaccard index. For the overall comparison, we performed statistical analysis using SPSS. Results: The radiological assessment of hemorrhage size follows the general trend of the quantitative measurements but does not reflect all short-term changes in volume. The semi-automatic method achieved the best volumetric and spatial overlap with manual segmentation but was also the most time consuming. The automatic method was the most time efficient but showed greater spatial discrepancies. The type of segmentation had a statistically significant effect on the Dice coefficient, Jaccard index and segmentation time. Discussion and conclusion: Segmentation performance depends on the method used as well as shape, size, distribution, homogeneity, contrast and time between the first occurence of hemorrhage and the imaging examination. Introduction of segmentation into clinical practice would significantly improve the clinical care of patients with traumatic brain injuries. As an initial approach, a two-stage combination of automatic method with user intervention would be the most appropriate.
|
