RESUMEN
BACKGROUND: Manually segmenting cardiac structures is time-consuming and produces variability in MRI assessments. Automated segmentation could solve this. However, current software is developed for adults without congenital heart defects (CHD). PURPOSE: To evaluate automated segmentation of left ventricle (LV) and right ventricle (RV) for pediatric MRI studies. STUDY TYPE: Retrospective comparative study. POPULATION: Twenty children per group of: healthy children, LV-CHD, tetralogy of Fallot (ToF), and univentricular CHD, aged 11.7 [8.9-16.0], 14.2 [10.6-15.7], 14.6 [11.6-16.4], and 12.2 [10.2-14.9] years, respectively. SEQUENCE/FIELD STRENGTH: Balanced steady-state free precession at 1.5 T. ASSESSMENT: Biventricular volumes and masses were calculated from a short-axis stack of images, which were segmented manually and using two fully automated software suites (Medis Suite 3.2, Medis, Leiden, the Netherlands and SuiteHeart 5.0, Neosoft LLC, Pewaukee, USA). Fully automated segmentations were manually adjusted to provide two further sets of segmentations. Fully automated and adjusted automated segmentation were compared to manual segmentation. Segmentation times and reproducibility for each method were assessed. STATISTICAL TESTS: Bland Altman analysis and intraclass correlation coefficients (ICC) were used to compare volumes and masses between methods. Postprocessing times were compared by paired t-tests. RESULTS: Fully automated methods provided good segmentation (ICC > 0.90 compared to manual segmentation) for the LV in the healthy and left-sided CHD groups (eg LV-EDV difference for healthy children 1.4 ± 11.5 mL, ICC: 0.97, for Medis and 3.0 ± 12.2 mL, ICC: 0.96 for SuiteHeart). Both automated methods gave larger errors (ICC: 0.62-0.94) for the RV in these populations, and for all structures in the ToF and univentricular CHD groups. Adjusted automated segmentation agreed well with manual segmentation (ICC: 0.71-1.00), improved reproducibility and reduced segmentation time in all patient groups, compared to manual segmentation. DATA CONCLUSION: Fully automated segmentation eliminates observer variability but may produce large errors compared to manual segmentation. Manual adjustments reduce these errors, improve reproducibility, and reduce postprocessing times compared to manual segmentation. Adjusted automated segmentation is reasonable in children with and without CHD. EVIDENCE LEVEL: 3. TECHNICAL EFFICACY: Stage 2.
