Automation of Wilms' tumor segmentation by artificial intelligence

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Oncology

Background:

• Manual segmentation of Wilms' tumors and kidneys for 3D reconstruction is time-consuming.
• Automated segmentation tools are needed to improve efficiency in pediatric oncology imaging.

Purpose of the Study:

• To develop an artificial intelligence tool for automated segmentation of pediatric Wilms' tumors and kidneys.
• To evaluate the performance of a novel CNN U-Net training method (OV²ASSION) for this task.

Main Methods:

• Manual segmentation of 14 CT scans by two experts.
• Automated segmentation using standard CNN U-Net and the OV²ASSION-trained CNN U-Net.
• Comparison of Dice indices for tumor and kidney segmentation accuracy.
• Estimation of time savings based on the number of manually segmented sections.

Main Results:

• Manual segmentation showed high inter-expert agreement (Dice index 0.95 for tumor, 0.87 for kidney).
• Fully automated CNN U-Net yielded suboptimal results (Dice index 0.69 for tumor, 0.27 for kidney).
• The OV²ASSION method achieved expert-level Dice indices (e.g., 0.97 for tumor, 0.94 for kidney with minimal manual input).

Conclusions:

• Fully automated segmentation of pediatric Wilms' tumors and kidneys remains challenging.
• The developed OV²ASSION training method for CNN U-Net enables expert-level precision.
• This innovative approach reduces expert intervention time by 80% while maintaining high segmentation accuracy.