Precision of automated cardiac chambers and great vessel volume segmentation in difficult cases using an open-source full body segmentation model

Zielsetzung To evaluate the precision of automated cardiac and great vessel segmentation using the ‘Total Segmentator’ (TS) by Wasserthal et al in patients with structural abnormalities.

Material und Methoden The study involved 89 CT scans, including a control group of clinical routine cases and 13 balanced study groups representing various structural abnormalities. Imaging was retrospectively chosen from multiple CT scanners from the years 2012 to 2024, utilizing soft tissue kernel and intravenous contrast agents. Automated segmentation of six mediastinal structures (left and right atrium, left and right ventricle, aorta, pulmonary artery) was followed by manual corrections. Statistical analysis with the dice coefficient, surface dice, jaccard index and volume similarity index followed to assess structural distinctiveness dependent on pathology or age.

Ergebnisse The model performed well (dice=0.9) on the clinical routine collective. It did not perform significantly (p<0.05) inferior on patients with spine and chest wall deformities (dice=0.87), with hypertrophic cardiomyopathy (dice=0.88), with aortic (dice=0.87), mitral (dice=0.88) and tricuspid valve replacements (dice=0.88). Patients with left ventricular assist devices were not segmented inferior in total (dice=0.87), but in the left atrium (surface dice=0.20) and the left ventricle (dice=0.69). In children age 0 the TS wasn’t able to detect the cardiac chambers and the great vessels (dice=0.23). The precision increased with age. Only the great arteries were segmented significantly inferior in patients between 10 and 16 years (total dice=0.88, aorta surface dice=0.51, pulmonary surface dice=0.37). In adults with congenital heart disease the algorithm performed inferior than on the clinical routine collective with a dice of 0.46 in patients with transposition of the great arteries, 0.38 in patients with a situs inversus and 0.53 in patients with Fontan circulation.

Schlussfolgerungen The algorithm provides dependable automatic results for segmentation of the heart chambers and the great arteries in clinical routine patients, in patients with spine and chest wall deformities and in replaced heart valves. In patients with a LVAD only the segmentation of the right heart and the great arteries is robust. Patients aged 0 to 9 need to manual segmentation. Adolescents between 10 and 16 years of age had good automatic results of the heart chambers, but the great arteries were limited. In adults with congenital heart disease the algorithm does not perform adequate.

Publication History

Article published online:
25 March 2025