Abstract:
Kidney cancer is a major health concern, with rising rates of diagnosis and mortal-
ity. Each year, thousands of people are diagnosed, and many lose their lives due to
late detection. Traditional diagnostic methods, while valuable, often fall short in
accuracy, leading to challenges in treatment planning and patient outcomes. While
computed tomography (CT) imaging is the gold standard for diagnosis, manual
tumor segmentation is time-consuming, prone to variability, and highly dependent
on radiologists’ expertise. Deep learning-based methods, particularly U-Net, have
shown a great promise in automating segmentation tasks. However, existing models
often struggle with ambiguous tumor boundaries, class imbalances, and misclassi-
fication of benign cysts. In this study, we implemented a U-Net Attention model
architecture, which integrates attention mechanisms into a U Net framework to en-
hance feature extraction, tumor localization, and segmentation accuracy of kidney
tumor segmentation from CT images. In the experiment, we follow a different ap-
proach in the pre-processing pipeline of our dataset. Our approach proves a powerful
way to segment kidney and tumor, leading to more accurate kidney disease diagno-
sis and treatment planning. We utilize the KiTS19 dataset for contrast-enhanced
CT images using semantic segmentation. Our model achieves a mean Dice score
of 0.85% and 0.70% for kidney and kidney tumors, respectively. It showcases the
potential to improve clinical kidney method decision-making.
