CT-Based Glioma Segmentation Using Deep Learning: Validation for Emergency Neuro-oncological Care.
Authors
Affiliations (5)
Affiliations (5)
- Department of Medical Physics, Institute of Health Sciences, Dokuz Eylul University, Izmir, Turkey.
- Department of Radiology, Faculty of Medicine, Dokuz Eylul University, Izmir, Turkey. [email protected].
- Department of Pathology, Faculty of Medicine, Dokuz Eylul Hospital, Izmir, Turkey.
- Department of Electrical and Electronics Engineering and Izmir Health Technologies Development and Accelerator (BioIzmir), Dokuz Eylul University, Izmir, Turkey.
- Department of Atomic and Molecular Physics, Dokuz Eylul University, Izmir, Turkey.
Abstract
Accurate and rapid delineation of diffuse gliomas is essential in emergency neuro-oncology, yet MRI is often unavailable. We present a deep-learning segmentation approach for routine non-contrast CT and evaluate its technical feasibility and WHO grade-stratified performance, demonstrating its potential as a decision-support tool in time-sensitive, resource-limited clinical settings. CT scans from 206 adults with histopathologically confirmed diffuse gliomas were retrospectively collected from a single center and were divided into a development cohort (n = 177) and an independent validation cohort (n = 29). The glioma segmentation network (GSN) consists of a ResNet-18 encoder and five-stage U-Net decoder, trained with inverse-frequency weighted cross-entropy to address class imbalance. Model development involved five-fold cross-validation on the development cohort (results reported in the main text). Primary outcome measures were Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95). Reference segmentations were generated by a board-certified neuroradiologist with over a decade of clinical experience using MITK. In the independent validation cohort, DSCs were 0.846 for grade 2 (95% CI 0.835-0.856), 0.806 for grade 3 (95% CI 0.800-0.811), and 0.802 for grade 4 (95% CI 0.796-0.809). Corresponding HD95 values were 13.677 mm (95% CI 12.70-14.65), 16.193 mm (95% CI 14.73-17.65), and 18.776 mm (95% CI 17.95-19.90). Inference throughput was 20 slices per second (95% CI 16.1-26.3). The proposed GSN demonstrates robust internal performance and clinically meaningful segmentation accuracy on independent CT data, supporting the technical feasibility of CT-based automated glioma delineation for emergency and resource-constrained settings. Evaluation was performed at the 2D slice level. Limitations include the single-center design and reliance on reference segmentations performed by a single neuroradiologist. Prospective multi-center validation is warranted before routine clinical implementation.