Uncertainty quantification in multi-parametric MRI-based meningioma radiotherapy target segmentation.

Area of Science:

• Medical imaging
• Artificial intelligence in radiology
• Neurosurgery

Background:

• Accurate segmentation of meningiomas on MRI is crucial for radiotherapy planning.
• Traditional U-Net models have limitations in quantifying segmentation uncertainty.

Purpose of the Study:

• To introduce and evaluate a spherical projection-based U-Net (SPU-Net) for improved meningioma segmentation.
• To enable uncertainty quantification in the segmentation process.

Main Methods:

• SPU-Net adapts U-Net by projecting 2D MRI scans onto a spherical surface to enhance details.
• Multiple projection centers generate diverse segmentations, with variance indicating uncertainty.
• Pixel-wise entropy calculations and Otsu's method quantify and aggregate uncertainty.

Main Results:

• SPU-Net achieved comparable segmentation performance to traditional U-Net (e.g., Dice coefficient 0.760 vs. 0.742).
• The model successfully quantified segmentation uncertainty, showing low uncertainty in accurate segments and high uncertainty at boundaries.
• Uncertainty maps revealed critical areas like GTV boundaries and dural tails.

Conclusions:

• SPU-Net offers a significant advantage by providing valuable uncertainty quantification alongside robust segmentation performance.
• This tool aids in identifying areas for potential manual correction, crucial for complex meningioma cases.
• SPU-Net represents an advanced and informative tool for neuro-oncology imaging analysis.