Dice Similarity Coefficient

In this study, we compared manual and automated segmentation algorithms in Computed Tomography (CT) imaging of the liver. Segmentation is the delineation of anatomical structures on a CT/MRI image for treatment and dose planning purposes. Accuracy in the segmentation of organs in CT imaging is a critical step in diagnosis and therapy planning. Normally, manual segmentation performed by trained practitioners is the gold standard for the delineation of these anatomical structures. However, manual segmentation is time-consuming, subject to interobserver differences, and also needs a level of expertise (Lee et al., 2024). Automated segmentation on the other hand, is a potential solution to the problems faced due to manual segmentation, despite this development, there is still an ongoing debate questioning the accuracy and reliability of automated segmentation due to little knowledge about its capabilities (Raudaschl et al., 2017). The main objectives of this study are to compare the accuracy and segmentation time of manual and automated segmentation, also to assess the inter-observer variability between 10 different observers. This study was carried out at the University of Benin Teaching Hospital (UBTH) using an anonymized CT dataset segmented manually using 3D-slicer and automatically using TotalSegmentator. A quantitative analysis of 100 segmentation masks produced by the segmentation of 10 anonymized CT images of patients by 10 observers (radiographers and radiotherapists) was conducted to evaluate the accuracy, time efficiency of automated algorithms and inter-observer variability of manual segmentation. The accuracy of automated segmentation was measured using the mean Dice Similarity Coefficient value, speed of both types of segmentation compared was measured using time and the inter-observer variability was measured using the Inter Class Correlation and Fleiss’ Kappa. The results obtained indicates that automated segmentation is as accurate as manual segmentation with a mean DSC value of 0.85±0.03 and also more time efficient with a 65% decrease in the segmentation time per image. This study also shows that the manual segmentations done by different observers are not far off from each other proving how reproducible it is. The Inter Class Correlation and Fleiss' Kappa was used to determine this with their score being 0.720 and 0.683 respectively. This indicates a moderate to strong agreement between segmentation done by different observers and if there is a strong agreement, the variability will be minimal. The DSC gotten is also greater than 0.7, giving a confirmation of how accurate
automated segmentation is. From the results above, it was concluded that automated segmentation is as accurate and faster than manual segmentation. It was also concluded that there are minimal differences between segmentations done by different observers. A wide clinical adoption, Training of radiographers, educational inclusion and further research is recommended to increase the awareness and integration of automated segmentation in various clinical systems.