Deep Learning-Based Glioma Segmentation of 2D Intraoperative Ultrasound Images: A Multicenter Study Using the Brain...

Area of Science:

• Medical imaging and artificial intelligence
• Neurosurgical oncology
• Image segmentation algorithms

Background:

• Intraoperative ultrasound (ioUS) offers real-time, portable, and cost-effective imaging in neurosurgery.
• Accurate segmentation of gliomas in ioUS images is crucial for enhanced interpretability but faces challenges like noise and artifacts.
• Existing methods struggle with anatomical variability and image quality issues inherent in ioUS.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) model for automated glioma segmentation in intraoperative ultrasound (ioUS) images.
• To assess the model's performance on a multicenter dataset, including external validation.
• To address limitations in ioUS image interpretation through advanced AI-driven segmentation.

Main Methods:

• Retrospective data collection from BraTioUS and ReMIND datasets for histologically confirmed gliomas.
• Training a CNN using the nnU-Net framework on B-mode ioUS images, with manual segmentation on the largest tumor slice.
• Stratified data splitting into training (70%) and testing (30%) subsets, with external validation on RESECT-SEG and Imperial College NHS Trust cohorts.

Main Results:

• The model achieved a median Dice Similarity Coefficient (DSC) of 0.90 on the hold-out test set.
• External validation demonstrated a DSC of 0.65 on the RESECT-SEG database and 0.93 on the Imperial-NHS cohort.
• Performance metrics (ASSD, HD95) indicated robust segmentation accuracy, particularly on the Imperial-NHS cohort.

Conclusions:

• Convolutional neural network-based glioma segmentation in intraoperative ultrasound is feasible across multiple institutions.
• The developed model shows promise for improving the interpretability and utility of ioUS in neurosurgical procedures.
• Future research should focus on enhancing segmentation detail and exploring real-time clinical integration for broader application.