Deep learning-based Wilms tumor segmentation to create 3D models for surgical planning: Implementation in the...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Surgical Planning

Background:

• Manual segmentation of Wilms tumor (WT) MRI for 3D models is time-consuming.
• Retrospective validation is common, but prospective clinical workflow evaluation is needed.
• Deep learning offers automated segmentation for improved surgical planning.

Purpose of the Study:

• To prospectively evaluate a deep learning-based WT segmentation method within a clinical workflow.
• To assess the feasibility of creating 3D models for surgical planning.
• To quantify the performance and time efficiency of automated segmentation.

Main Methods:

• Developed and retrospectively validated a nnU-Net based deep learning segmentation method.
• Implemented the method in a clinical workflow for prospective testing on 10 WT patients.
• Quantified performance using Dice scores and analyzed segmentation error types and correction times.

Main Results:

• Automated segmentation was sufficient in 20% of patients; 80% required corrections.
• Median correction time was 11 minutes, with most corrections under 15 minutes.
• High Dice scores achieved for kidney (1.00) and tumor (0.98), with common errors including under-segmentation and border misidentification.

Conclusions:

• Automated WT segmentation for 3D model creation is feasible in clinical settings.
• The deep learning method significantly reduces manual effort in surgical planning.
• This approach enhances efficiency and accuracy in preparing for Wilms tumor surgery.