gms | German Medical Science

Introduction: Segmentation of medical image data achieves increasing importance. In the field of ENT-surgery segmentation of paranasal sinuses is not yet used in everyday surgical workflow. As it seems to be very accurate manual segmentation is often used as a reference for the evaluation of automatic segmentation. The purpose of this study is to describe the inter- and intraindividual variability of manual segmentation.

Methods: In a first experiment ten ENT- surgeons and ten medical students outlined the right maxillary sinus and the right ethmoid sinuses manually. In a second experiment two experienced participants outlined the same sinuses five times consecutively. The segmentation procedure was accomplished with custom software. To evaluate the results we used three indices: volume, extension and visual analysis.

Results: Our first experiment showed the interindividual variability of manual segmentation. It was higher for the ethmoidal sinuses than for the maxillary sinuses. The second experiment showed that the intraindividual variability was lower than the interindividual.

Discussion: There are multiple reasons for the variability. It can be caused by different levels of experience and accuracy as well as different interpretation of the CT data. Variability may be particularly high in paranasal sinuses due to their complex anatomy.

The inter- and intraindividual variability show that the segmentation result of one manual segmenter is not an adequate gold standard for the evaluation of automatic segmentation algorithms. Standardized and reproducible segmentation results are absolutely necessary for medical applications. For the evaluation of medical image segmentation reference images based on multiple manual segmentation results should be used.