A multi-paradigm evaluation spanning pixels to voxels for deep learning-based kidney tumor segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Automated kidney tumor segmentation from CT scans is vital for renal cell carcinoma (RCC) management.
• Deep learning models show high accuracy (Dice scores >0.97) but often fail in clinical settings due to false positives and computational limits.

Purpose of the Study:

• To bridge the gap between high segmentation accuracy and clinical applicability of deep learning models for kidney tumors.
• To systematically evaluate diverse deep learning architectures for their clinical utility beyond quantitative metrics.

Main Methods:

• Evaluation of six 2D and 3D deep learning architectures (U-Net, MedSAM, nnU-Net, UNETR, Total Segmentator, MIScnn) on the KiTS19 dataset.
• Emphasis on false positive analysis, boundary accuracy, and computational feasibility (VRAM, GPU requirements).

Main Results:

• MONAI U-Net (Dice: 0.98) had excessive false positives. nnU-Net (Dice: 0.82) offered balanced performance but required 16GB VRAM. MedSAM (Dice: 0.99) achieved high accuracy with few false positives but needed high-end GPUs.
• UNETR training was computationally constrained. High Dice scores did not directly correlate with clinical applicability.

Conclusions:

• Clinical utility of automated kidney tumor segmentation requires careful consideration of false positives and computational resources.
• Actionable insights are provided for developing clinically feasible segmentation tools for renal oncology, aiding treatment planning and monitoring.