Abstract
BACKGROUND: This study evaluates the contouring variability among observers using MR images reconstructed by different sequences and quantifies the differences of automatic segmentation models for different sequences. PATIENTS AND METHODS: Eighty-three patients with pelvic tumors underwent T1-weighted image (T1WI), contrast enhanced Dixon T1-weighted (T1dixonc), and T2-weighted image (T2WI) MR imaging on a simulator. Two observers performed manual delineation of the bladder, anal canal, rectum, and femoral heads on all images. Contour differences were used to analyze the interobserver and intersequence variability. A single-sequence automatic segmentation network was established using the U-Net network, and the segmentation results were analyzed. RESULTS: Variability analysis among observers showed that the bladder, rectum, and left femoral head on T1WI yielded the highest dice similarity coefficient (DSC) and the lowest 95% Hausdorff distance (HD) (all three sequences). Regarding sequence variability analysis for the same observer, the difference between T1WI and T2WI was the smallest. The DSC of the bladder, rectum, and femoral heads exceeded 0.88 for T1WI-T2WI. The differences between automatic segmentations and manual delineations were minimal on T2WI. The averaged DSC of automatic and manual segmentation of all organs on T2WI exceeded 0.81, and the averaged 95% HD value was lower than 7 mm. Similarly, the sequence variability analysis of automatic segmentation indicates that the automatic segmentation differences between T2WI and T1WI are minimal. CONCLUSIONS: T1WI and T2WI yielded better results in manual delineation and automatic segmentation, respectively. The analysis of variability among three sequences indicates that the yielded good similarity outcomes between the T1WI and T2WI cases in manual and automatic segmentation. We infer that the T1WI and T2WI (or their combination) can be used for MR-only radiation therapy.