Abstract
Objectives: To investigate the efficiency and performance of artificial intelligence (AI)-assisted prostate segmentation by radiographers compared to manual segmentation by radiologists and standalone AI.
Materials & Methods: T2-weighted transverse MRI images from two publicly available datasets, PROMISE12 and ProstateX and manual segmentations by two groups of radiologists, were used. Three radiographers from the same country segmented the prostate manually and with assistance from a 3D nnU-Net model trained on the PROMISE12 data. Dice Similarity Coefficient (DSC), Hausdorff Distance (HD95), Relative Volume Difference (RVD), and time consumption were compared using the Wilcoxon signed-rank test in the ProstateX data, with one group radiologists’ segmentation as reference standard. Intraclass Correlation Coefficient (ICC) was used to evaluate inter-rater variability of segmented prostate volumes.
Results: AI-assisted radiographer segmentation showed comparable performance to manual segmentation by radiologists (median DSC 0.894) with significantly reduced time consumption compared to manual radiographer segmentation (60 vs. 215.5 seconds, p<0.001). Inter-rater variability among radiographers was eliminated with AI assistance (ICC 1.00, p<0.001).
Conclusion: AI-assisted radiographer prostate segmentation is a viable alternative to manual segmentation by radiologists, offering improved efficiency and consistency compared to manual segmentation by radiographers. Further prospective multi-center validation is needed to evaluate its impact on clinical prostate cancer imaging workflows.
