Atlas Based Automatic Segmentation, A Clinical Implementation & Assessment of Accuracy
Presentations
TU-I345-IePD-F3-3 (Tuesday, 7/12/2022) 3:45 PM - 4:15 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 3
Purpose: Manually drawing contours on a patients CT is a time-consuming process. In addition, there is also a significant amount of intra- and inter-observer inconsistencies during the contouring procedure: the different clinicians may interpret boundaries of each organ differently. A solution to these two hurdles incorporates the implementation of an atlas-based automatic contouring software. The accuracy of such an approach has to be evaluated to determine if the software is able to produce contours similar to those manually drawn, which are used as the “gold standard” contours. Atlas based contour, in contrast to a cloud based approach, allows for auto-segmentation to be customized to a clinics style of contouring, while still potentially improving consistency.
Methods: Atlases were created in 5 sites of the body using MIM then applied and compared to previously treated and clinically accepted contours on patient CTs. The individual organs of both atlas-based contours and manually drawn, clinically accepted contours were compared volumetrically with the Dice Similarity Coefficient (DSC).100 CTs were used to created atlases based on 5 different anatomical regions: Pelvic, Abdomen, Chest, Head and Neck, and Brain. The SC was then calculated to compare the accuracy of the atlas-based auto-segmtations. Testing was carried out on a further 20 patients per site.
Results: The average DSC per site are Abdomen 0.60, Head and Neck 0.69, Pelvis 0.65, Thoracic 0.81, Brain 0.64. Smaller contours tended to have smaller (worse) DSC due to the metric being a ratio of volumes and does not indicate that that they would be clinically unacceptable.
Conclusion: Atlas based auto-segmentation was successfully implemented into our clinic. DSC vales were the highest for the brain (0.98), liver (0.81), femoral heads (0.80), brainstem (0.78) and mandible (0.77). The worst structures were larynx (0.47), parotids (0.41), stomach (0.40), cochlea (037) and optic chiasm (0.21).
Keywords
Taxonomy
IM/TH- Image Segmentation Techniques: Model and Atlas-based
Contact Email