INTRODUCTION
Computed tomography (CT) is an indispensable imaging modality in urological practice and provides critical diagnostic information for renal tumors, cysts, and structural abnormalities [1, 2]. However, cumulative radiation exposure from repeated CT examinations remains an important patient-safety concern, particularly for patients who require serial imaging surveillance [1, 3].
Recent advances in artificial intelligence (AI), particularly deep learning-based segmentation algorithms, have demonstrated strong performance in automated kidney delineation on CT images [4-6]. These AI tools have the potential to improve diagnostic workflow efficiency and consistency [2, 7-9]. Nevertheless, their performance under suboptimal imaging conditions, particularly reduced radiation dose settings, remains insufficiently characterized.
The principle of “as low as reasonably achievable” requires radiation exposure to be minimized while maintaining diagnostic image quality [3]. As AI-assisted analysis becomes increasingly integrated into clinical workflows, understanding how dose reduction affects AI performance is essential for establishing evidence-based dose optimization protocols.
Several studies have examined the effects of low-dose CT on image quality and diagnostic accuracy [10-12]. However, systematic analyses of the relationship between radiation dose reduction and deep learning-based segmentation performance in renal imaging remain scarce [13]. Previous research has primarily focused on image quality metrics or human reader performance and has not fully addressed the distinct sensitivity profile of AI algorithms to noise-related image degradation.
This study aimed to quantitatively evaluate the effect of simulated radiation dose reduction on deep learning-based renal segmentation performance using the 2021 Kidney and Kidney Tumor Segmentation Challenge (KiTS21) dataset [5]. Specifically, this study sought to (1) evaluate segmentation performance across 4 dose levels (100%, 50%, 25%, and 10%), (2) identify the dose threshold below which performance degradation becomes clinically significant, and (3) provide evidencebased recommendations for dose optimization in AI-assisted renal imaging.
MATERIALS AND METHODS
Dataset
This study used the KiTS21 dataset, a publicly available collection of contrast-enhanced CT volumes with expert segmentation labels [5]. The dataset comprises 300 cases with voxel-level annotations generated by majority voting among 3 independent expert annotators. Segmentation labels include kidney parenchyma, tumor, and cyst. For this study, all nonbackground labels were merged into a single kidney class for binary segmentation of background versus kidney tissue, allowing the analysis to focus on renal delineation. After quality filtering, 299 cases comprising 22,865 axial slices containing kidney tissue were included. Because the KiTS21 dataset is publicly available and was collected with prior institutional ethics approval, this secondary analysis was deemed exempt from additional review.
Dose Simulation
Radiation dose reduction was simulated using a physics-based Poisson noise model. CT image noise follows a Poisson distribution, in which reducing the dose by a factor of k increases noise variance proportionally [11, 12]. Standard CT windowing with a center of 40 Hounsfield units (HU) and width of 400 HU was applied to all images before noise addition. Four dose levels were generated: 100% dose, corresponding to the original images without added noise, and simulated 50%, 25%, and 10% doses. Noise simulation was applied uniformly to all axial slices, generating 4 complete imaging datasets from the same underlying anatomy (Figs. 1 and 2).
Model Architecture and Training
A 2-dimensional (2D) U-Net architecture [14] with a ResNet34 encoder [15] pretrained on ImageNet was used for renal segmentation. The model was implemented using the segmentation-models-pytorch library [16]. Training was performed exclusively on standard-dose images to simulate a realistic clinical scenario in which the model encounters imaging conditions that differ from those used during training.
Training parameters included an input size of 256×256 pixels with 3-channel replication for compatibility with the pretrained encoder, a combined loss function with equal weighting of Dice loss and cross-entropy loss, the Adam optimizer with a learning rate of 1×10-4and weight decay of 1×10-5, learningrate reduction on plateau with a factor of 0.5 and patience of 7 epochs, early stopping with patience of 15 epochs, and a maximum of 100 training epochs. Training data augmentation included random horizontal and vertical flipping and rotation.
Five-fold cross-validation was performed with patient-level stratification to prevent data leakage between the training and test sets. In each fold, approximately 20% of cases were reserved for testing.
Performance Evaluation
Case-level evaluation was performed to minimize the influence of outlier slices, such as slices with minimal kidney representation at the superior or inferior poles. For each case, slice-level metrics were first averaged to generate a single per-case value, and subsequent statistical analyses were conducted across cases.
Four evaluation metrics were calculated: (1) the Dice similarity coefficient (DSC) [17], which measures volumetric overlap between predicted and ground-truth segmentations; (2) intersection over union (IoU), or the Jaccard index; (3) 95th-percentile Hausdorff distance (HD95) [18], which quantifies boundary accuracy in pixels; and (4) volumetric error, calculated as the absolute percentage difference between predicted and ground-truth kidney volumes. Slices without kidney tissue in the ground truth were excluded from metric calculation to avoid artificially inflating performance.
Statistical Analysis
The Wilcoxon signed-rank test was used for paired comparisons between the standard dose and each reduced dose level (50%, 25%, and 10%), because the same cases were evaluated under all dose conditions. Effect size was calculated as r=|Z|/√N, where N is the number of paired cases. A P-value <0.05 was considered statistically significant, and Bonferroni correction was applied for multiple comparisons. All statistical analyses were performed using SciPy ver. 1.11 (SciPy developers, https://scipy.org/).
RESULTS
Segmentation Performance Across Dose Levels
Table 1 summarizes case-level segmentation performance across the 4 simulated dose levels. At the standard dose, the model achieved a mean DSC of 0.948±0.044 and an IoU of 0.918± 0.054. Performance declined gradually and monotonically as the dose decreased: the DSC was 0.945±0.046 at 50% dose, 0.939±0.052 at 25% dose, and 0.921±0.069 at 10% dose (Fig. 3A and B). The total DSC reduction from 100% to 10% dose was approximately 0.027, corresponding to a 2.8% relative decrease. Visual comparisons of segmentation outputs across dose levels are presented in Fig. 4.
Statistical Analysis
The Wilcoxon signed-rank test showed statistically significant differences in DSC between the standard dose and all reduced dose levels (Table 2). All comparisons were statistically significant (P<0.001). Effect-size analysis indicated that these differences were not merely artifacts of the large sample size: effect sizes increased progressively from r=0.591 for 100% versus 50% dose to r=0.742 for 100% versus 25% dose and r=0.845 for 100% versus 10% dose. All effect sizes were classified as large (r>0.5). This progressive increase suggests that dose reduction produced increasingly substantial and clinically mean-ingful segmentation-performance degradation.
Dose-Performance Relationship
The dose-performance relationship showed a nonlinear pattern. Between 100% and 50% dose, performance decline was minimal, with an approximate DSC decrease of 0.003, suggesting substantial model robustness to moderate noise increases. Between 50% and 25% dose, a slightly larger decline was observed, with an approximate DSC decrease of 0.006. The most pronounced degradation occurred between 25% and 10% dose, with an approximate DSC decrease of 0.018, indicating accelerated performance loss at extreme dose-reduction levels.
HD95 followed a similar pattern, increasing from 4.73 pixels at 100% dose to 6.58 pixels at 10% dose, corresponding to a 39% increase. Volumetric error increased markedly at 10% dose (12.64% vs. 8.94% at 100% dose), reflecting greater segmentation-boundary variability under severe noise conditions (Fig. 5).
DISCUSSION
This study systematically evaluated the effect of simulated radiation dose reduction on deep learning-based renal segmentation performance. The key finding was that a U-Net model trained exclusively on standard-dose CT images maintained clinically acceptable segmentation accuracy, defined as DSC >0.93, down to 25% of the standard dose, with substantial performance preservation even at 50% dose reduction. At 10% of the dose, DSC declined more substantially to 0.921, identifying a practical lower bound for dose optimization in this simulation setting.
The observed robustness of renal segmentation to dose reduction may be attributable to several factors. First, the kidney is a relatively large, well-defined anatomical structure with high contrast against surrounding tissues, particularly on contrastenhanced CT. The distinct intensity profile of renal parenchyma may provide sufficient signal-to-noise ratio even under reduced-dose conditions. Second, the U-Net architecture with pretrained encoder features may have learned hierarchical representations that are inherently noise tolerant, because ImageNet pretraining exposes the model to natural image variations that may share some statistical properties with imaging noise.
The case-level evaluation methodology merits discussion. Initial slice-level analysis produced misleading results, with a mean DSC of 0.947 and a standard deviation of 0.133, yielding error bars that obscured dose-dependent differences. This high variance was driven by outlier slices at the kidney poles, where only a few pixels of kidney tissue were present and DSC calculations were therefore unstable. By aggregating slices to the case level before analysis, the standard deviation decreased substantially to approximately 0.044, producing clearer dose-response curves. This finding underscores the importance of selecting an appropriate evaluation granularity in segmentation studies.
The nonlinear dose-performance relationship observed in this study has practical implications. The minimal performance decline between 100% and 50% dose suggests that moderate dose reduction may be feasible without a meaningful effect on AI-assisted renal segmentation. The accelerating decline below 25% dose suggests a potential inflection point, below which image noise begins to more substantially affect the model’s ability to delineate kidney boundaries. This information may inform dose optimization protocols in clinical settings where AI-assisted analysis is used.
From a clinical perspective, acceptable performance at 25% dose has practical implications for renal mass surveillance, in which repeated imaging could benefit from up to 75% cumulative dose reduction while preserving AI-assisted volumetric accuracy. This finding is particularly relevant for pediatric patients, who have greater radiation sensitivity, and for older adults with impaired renal function, who may benefit from low-dose noncontrast CT combined with AI segmentation. However, clinical implementation requires validation using actual low-dose acquisitions, because simulated noise may not fully replicate real imaging conditions.
Comparison with the literature supports these findings. The baseline DSC of 0.948 at standard dose is consistent with reported U-Net performance on the KiTS21 dataset, including an nnU-Net benchmark DSC of 0.97 for renal segmentation [2], particularly given that the present model used a simpler 2D architecture without extensive preprocessing. The relative robustness to noise is also consistent with previous studies reporting that convolutional neural networks can exhibit inherent noise tolerance in medical image analysis tasks [11, 13].
This study used a standard-dose-only training paradigm, representing a conservative but clinically realistic scenario. Alternative strategies, such as mixed-dose training or fine-tuning on a small subset of low-dose images, could enable the model to learn dose-invariant features and potentially recover performance at extreme dose reductions. The standard-dose-only results therefore provide an important baseline against which dose-adaptive strategies can be benchmarked.
The generalizability of these findings to other anatomical structures also warrants consideration. The kidney’s relatively large size and high tissue contrast on contrast-enhanced CT inherently favor noise-tolerant segmentation. Other large abdominal organs, such as the liver and spleen, may show comparable robustness. In contrast, smaller structures with lower contrast, such as adrenal glands, lymph nodes, or small renal tumors, are likely more susceptible to noise-induced degradation. Multiorgan comparative studies could help define organ-specific dose optimization thresholds.
Several limitations should be acknowledged. First, although the Poisson noise simulation was physically motivated, it may not fully capture the complexity of actual low-dose CT imaging, which involves additional factors such as electronic noise, beam-hardening artifacts, and iterative reconstruction algorithms [10]. Second, this study focused exclusively on renal segmentation; performance degradation patterns for smaller or less well-defined structures, such as renal tumors and cysts, may differ substantially. Third, the use of a single model architecture, a 2D U-Net, limits the generalizability of the findings; more advanced architectures, such as 3-dimensional U-Net, nnU-Net, and transformers, may exhibit different noise-sensitivity profiles [6, 19]. Fourth, this study used simulated rather than actual clinical low-dose CT data, and validation with real low-dose acquisitions would strengthen the conclusions.
Future studies should address these limitations by validating the findings with paired low-dose and standard-dose clinical CT acquisitions, extending the analysis to tumor and cyst segmentation, evaluating multiple model architectures for comparative robustness, and investigating noise-aware training strategies, such as training with mixed dose levels, as a potential approach to improving robustness.
In conclusion, deep learning-based renal segmentation using a 2D U-Net trained on standard-dose CT images demonstrated substantial robustness to simulated radiation dose reduction. Case-level evaluation across 299 cases showed that segmentation performance remained clinically acceptable, with a DSC>0.93, down to 25% of the standard dose. A clinically relevant performance threshold was identified around the 10% dose level, where accelerated performance degradation was observed. These findings provide quantitative evidence supporting dose optimization strategies in AI-assisted renal CT imaging and suggest that moderate dose reduction may be implemented without critically compromising automated renal segmentation accuracy, pending validation in real low-dose clinical acquisitions.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Print
