Kidney Tumor Segmentation With a Multistage Adaptive Boundary-Aware Network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate kidney tumor segmentation is crucial for effective surgical planning.
• Challenges include indistinct tumor boundaries and morphological variability in CT images.

Purpose of the Study:

• To develop an advanced deep learning model for precise kidney tumor segmentation.
• To improve accuracy and boundary detection in CT scans for better surgical planning.

Main Methods:

• Proposed the Adaptive Boundary-Aware Network (MABS-Net) incorporating boundary-aware multiscale feature extraction.
• Implemented an adaptive three-stage cascaded strategy for progressive refinement.
• Utilized contrastive learning with online hard example mining for enhanced feature discrimination.

Main Results:

• MABS-Net achieved a Dice coefficient of 0.891 ± 0.034 on the KiTS19 dataset, outperforming the nnU-Net baseline.
• Reduced 95% Hausdorff distance (HD95) to 6.73 ± 2.28 mm and improved boundary Dice score by 5.8%.
• Demonstrated computational efficiency (0.53 s/case) and provided pixel-wise uncertainty maps.

Conclusions:

• MABS-Net offers a superior solution for automated renal tumor segmentation, enhancing accuracy and boundary definition.
• The model's boundary-aware design and uncertainty quantification support reliable clinical decision-making.
• MABS-Net presents a promising tool for improving kidney tumor analysis and surgical planning.