INTRODUCTION

Computational methods for segmentation of clinical MR imaging are advancing due to more affordable machine learning algorithms. However, accurately delineating upper abdomen organs remains challenging. This study compares MRISegmentator and TotalSegmentator in processing Dixon-based MR Images.

MATERIALS AND METHODS

A cohort of 29 subjects (Mean Age: 62±15, 16 F, 11 with liver-related diseases) was included. The imaging protocol included m-Dixon Quant along with conventional T1-weighted Dixon imaging at 3T. The liver was manually segmented in 3DSlicer by a trained operator, serving as reference standard. Automatic outputs from MRI- and TotalSegmentator were evaluated by comparing segmented volumes and spatial overlap using the Dice Similarity Coefficient (DSC).

RESULTS

Both automatic algorithms overestimated liver volumes, 6.92% for MRISegmentator and 11.99% for TotalSegmentator, relative to manual segmentation; however, liver segmentation between both algorithms achieved optimal spatial agreement (mean DSC = 0.93±0.049). In contrast, the same comparison applied to the pancreas results in a mean DSC = 0.51±0.21.

Automated segmentation of the liver resulted in a mean DSC = 0.58±0.08 compared to manual segmentation.

CONCLUSION

Automated segmentation of the upper abdomen performs well for the liver but yields suboptimal results for the pancreas. This is likely due to the pancreas’s complex structure, low imaging contrast, partial voluming, and the potential presence of artifacts in the scans. Although retraining the models specifically for the pancreas might improve their performance, there is a risk that such focused retraining could reduce the models’ ability to generalize over more common contrasts.