Impact of Simulated Radiation Dose Reduction on Deep Learning-Based Renal Segmentation Performance: A Simulation...

This study aimed to quantitatively evaluate the effect of simulated radiation dose reduction on deep learning-based renal segmentation performance and to identify a clinically acceptable minimum dose threshold.

Using the KiTS21 (2021 Kidney and Kidney Tumor Segmentation Challenge) dataset, which included 299 contrastenhanced computed tomography volumes with expert segmentation labels, 4 dose levels were simulated: 100%, 50%, 25%, and 10%. Dose reduction was simulated using Poisson noise modeling. A 2-dimensional U-Net with a ResNet34 encoder was trained exclusively on standard-dose images and evaluated across all dose levels using 5-fold cross-validation. Case-level performance was assessed using the Dice similarity coefficient (DSC), intersection over union, 95th-percentile Hausdorff distance (HD95), and volumetric error. Statistical significance was evaluated using the Wilcoxon signed-rank test with effect-size analysis.

At the standard dose, the model achieved a case-level DSC of 0.948±0.044. Performance remained stable at 50% dose (0.945±0.046), declined moderately at 25% dose (0.939±0.052), and decreased more substantially at 10% dose (0.921±0.069). The Wilcoxon signed-rank test showed statistically significant differences between 100% dose and all reduced dose levels (P<0.001). HD95 increased from 4.73±4.82 pixels at 100% dose to 6.58±6.74 pixels at 10% dose.

Deep learning-based renal segmentation demonstrated substantial robustness to simulated dose reduction. Performance remained clinically acceptable, with a DSC>0.93, down to 25% of the standard dose, suggesting that substantial dose reduction may be feasible without critically compromising artificial intelligence-assisted renal segmentation. The marked performance decline at 10% dose identifies a potential lower bound for clinical dose optimization.