tsta

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Benchmarking nnU-Net vs. Custom 3D U-Net for Kidney Tumor Segmentation: A Controlled Study on KiTS19 Dataset

Authors :

Ariya Soleimany¹ Masoud Noroozi² Mohammad Saber Azimi³ Alireza Karimian⁴ Jafar Majidpour⁵ Hossein Arabi⁶

1- University of Isfahan 2- University of Isfahan 3- Shahid Beheshti University 4- University of Isfahan 5- University of Raparin 6- Geneva University Hospital

Keywords :

deep learning،medical image segmentation،3D U-Net،nnU-Net،kidney tumor

Abstract :

This study presents a controlled comparison between nnU-Net and a custom three-dimensional U-Net architecture for kidney and tumor segmentation using the KiTS19 dataset. Both models were trained under identical conditions, using the same preprocessing, data augmentation, and training protocols to isolate the effects of architecture and optimization strategies. Evaluation was conducted on 210 labeled cases, with 140 used for training and 70 for testing, using the Dice similarity coefficient and intersection over union metrics. The nnU-Net achieved Dice scores of 0.9683 for kidney and 0.8166 for tumor, while the custom U-Net obtained scores of 0.9450 and 0.6296, respectively. Similarly, intersection over union scores were 0.9390 and 0.7176 for nnU-Net, compared to 0.8997 and 0.5167 for the custom model. A non-parametric Wilcoxon signed-rank test confirmed that these differences were statistically significant, with a p-value that was extremely small and very close to zero. These findings highlight the role of deep supervision, architectural scaling, and automated configuration in enhancing segmentation accuracy for challenging medical imaging tasks.