Related Experiment Video
Updated: Apr 15, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.7K
CT-Based Glioma Segmentation Using Deep Learning: Validation for Emergency Neuro-oncological Care
Zohal Alnour Ahmed Emam1, Emel Ada2, Berrin Çavuşoğlu1
1Department of Medical Physics, Institute of Health Sciences, Dokuz Eylul University, Izmir, Turkey.
Journal of Imaging Informatics in Medicine
|April 13, 2026
Summary
A new deep learning model accurately segments diffuse gliomas on CT scans, offering a vital tool for emergency neuro-oncology when MRI is unavailable. This glioma segmentation network (GSN) shows promising results for resource-limited settings.
Area of Science:
- Neuro-oncology
- Medical Imaging
- Artificial Intelligence
Background:
- Accurate diffuse glioma delineation is critical in emergency neuro-oncology.
- Magnetic Resonance Imaging (MRI) is often unavailable in time-sensitive situations.
- Non-contrast Computed Tomography (CT) is a more accessible imaging modality.
Purpose of the Study:
- To develop and evaluate a deep learning segmentation approach for diffuse gliomas using routine non-contrast CT.
- To assess the technical feasibility and WHO grade-stratified performance of the model.
- To demonstrate its potential as a decision-support tool in resource-limited clinical settings.
Main Methods:
- A retrospective collection of 206 adult diffuse glioma CT scans from a single center.
- Development of a glioma segmentation network (GSN) using a ResNet-18 encoder and U-Net decoder.
- Training with inverse-frequency weighted cross-entropy and five-fold cross-validation on a development cohort (n=177) and validation on an independent cohort (n=29).
Main Results:
- The GSN achieved Dice Similarity Coefficients (DSC) of 0.846 (grade 2), 0.806 (grade 3), and 0.802 (grade 4) in the independent validation cohort.
- Corresponding 95th percentile Hausdorff Distance (HD95) values were 13.677 mm, 16.193 mm, and 18.776 mm, respectively.
- Inference throughput reached 20 slices per second, demonstrating rapid processing.
Conclusions:
- The proposed GSN demonstrates robust performance and clinically meaningful segmentation accuracy on independent CT data.
- This supports the technical feasibility of CT-based automated glioma delineation for emergency and resource-constrained settings.
- Prospective multi-center validation is recommended before routine clinical implementation.
