Can Urol Assoc J. 2026 Feb 13. doi: 10.5489/cuaj.9476. Online ahead of print.
ABSTRACT
INTRODUCTION: Automated segmentation using artificial intelligence (AI) has the potential to rapidly perform three-dimensional (3-D) segmentation of small renal masses (SRM). The objective of this study was to test for clinically and statistically significant differences in time spent segmenting, accuracy, and reliability when comparing manual and automated segmentation of computed tomography (CT) scans with SRM.
METHODS: Patients with a CT scan, SRM <4 cm, and renal neoplasm were identified through an institutional database. Of the 854 patients identified, 184 were excluded. Forty test cases were randomly selected. There were 630 cases for training (using nnU-Net) to which 488 cases from the KiTS23 open-source data set were added. Each of the test cases was segmented by a radiologist, a urologist, and the AI model. Time to segment and Dice coefficients were compared. Deidentified segmented CTs were provided to two independent radiologists who attempted to identify the segmentor and rated the acceptability of the segmented images on a five-point Likert scale.
RESULTS: There were 39 cases with complete timing data. The median time for the AI model to segment was one third of the radiologist’s (152.4 s, interquartile range [IQR] 120.9-177.8 vs. 450 s, IQR 318.8-551.2) and about one-fifth of the urologist’s (800.0 s, IQR 492.0-1538.0) (p<0.001). There was a high degree of inter-rater reliability (median Dice coefficients 0.86-0.90, p=0.09). The scoring radiologists were able to correctly identify the true segmentor in 61.6% of cases (p <0.001). The AI segmentations were scored highest among the three segmentors (median score 4.1/5, standard deviation [SD] 1.0) compared to 3.8 (SD 0.7) for the radiologist, and 3.3 (SD 0.7) for the urologist.
CONCLUSIONS: Automated segmentation of CT scans for patients with SRM was efficient, accurate, and acceptable in this study. This approach has the potential to greatly improve the clinical use of radiomics to assess medical images for these patients.
PMID:41701890 | DOI:10.5489/cuaj.9476
