U-ConvNext: A Robust Approach to Glioma Segmentation in Intraoperative Ultrasound.

Area of Science:

• Neurosurgery
• Medical Imaging
• Artificial Intelligence

Background:

• Intraoperative tumor imaging is vital for safe neurosurgical resection, particularly for low-grade gliomas.
• Ultrasound offers convenience but faces limitations in manual tumor segmentation accuracy and time.
• Accurate segmentation is crucial for maximizing tumor removal while preserving healthy brain tissue.

Purpose of the Study:

• To develop and evaluate a learning-based model for precise low-grade glioma segmentation in intraoperative ultrasound images.
• To address the limitations of manual segmentation and enhance tumor detection during neurosurgery.
• To provide reliable uncertainty quantification for improved surgical decision-making.

Main Methods:

• A novel U-net-based architecture (U-ConvNext) integrating ConvNeXt V2 blocks, global response normalization, and inception layers was developed.
• CutMix data augmentation was employed for enhanced texture detection in semantic segmentation.
• Conformal segmentation, a new conformal prediction method, was introduced for uncertainty quantification.

Main Results:

• The U-ConvNext model achieved high hold-out test Dice scores (84.63%, 74.52%, 90.82%) on RESECT dataset subsets.
• Performance demonstrated significant improvements (13-31%) over state-of-the-art methods.
• External validation on the ReMIND dataset showed robust performance (Dice score 79.17%) with minimal calibration error increase.

Conclusions:

• The proposed U-ConvNext model offers a significant advancement in automated low-grade glioma segmentation using ultrasound.
• The integration of architectural innovations, data augmentation, and uncertainty quantification enhances segmentation accuracy and reliability.
• This AI-driven approach holds promise for improving maximal safe resection rates in neurosurgery.