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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning models are increasingly used for medical image analysis, including renal segmentation.
• Radiation dose reduction in computed tomography (CT) is crucial for patient safety but can impact image quality and diagnostic accuracy.
• Evaluating the performance of AI algorithms under reduced radiation doses is essential for clinical implementation.

Purpose of the Study:

• To quantitatively assess the impact of simulated radiation dose reduction on deep learning-based renal segmentation.
• To determine a clinically acceptable minimum radiation dose threshold for AI-assisted renal segmentation.

Main Methods:

• Utilized the KiTS21 dataset (299 contrast-enhanced CT volumes) with expert segmentation labels.
• Simulated four dose levels: 100%, 50%, 25%, and 10% of the standard dose using Poisson noise modeling.
• Trained a 2D U-Net with ResNet34 encoder on standard-dose images and evaluated performance across all dose levels using 5-fold cross-validation, assessing Dice Similarity Coefficient (DSC), Intersection over Union, Hausdorff distance (HD95), and volumetric error.

Main Results:

• The model achieved a high DSC of 0.948±0.044 at standard dose.
• Performance remained stable at 50% dose (DSC 0.945±0.046) and declined moderately at 25% dose (DSC 0.939±0.052).
• A significant performance decrease was observed at 10% dose (DSC 0.921±0.069), with HD95 increasing from 4.73±4.82 to 6.58±6.74 pixels.

Conclusions:

• Deep learning-based renal segmentation demonstrates considerable robustness to simulated radiation dose reduction.
• Clinical acceptability (DSC > 0.93) was maintained down to 25% of the standard dose, suggesting feasibility of substantial dose reduction.
• The substantial performance decline at 10% dose highlights a potential lower bound for dose optimization in AI-assisted renal segmentation.