Enhanced glioma semantic segmentation using U-net and pre-trained backbone U-net architectures

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Gliomas exhibit distinct sub-regions (edema, necrotic, active tumor) crucial for treatment planning.
• Accurate segmentation of these regions is vital for effective glioma management.

Purpose of the Study:

• To evaluate U-Net and pre-trained U-Net backbone networks for glioma semantic segmentation using multimodal MRI data.
• To assess the performance of ResNet, Inception, and VGG as backbones within the U-Net architecture.

Main Methods:

• Utilized the BraTS 2021 challenge dataset for training, validation, and testing.
• Applied U-Net architecture with ImageNet pre-trained backbones (ResNet, Inception, VGG).
• Assessed network performance using Accuracy (ACC) and Intersection over Union (IoU).

Main Results:

• ResNet-U-Net trained with T1 post-contrast enhancement (T1Gd) demonstrated the highest ACC and IoU for necrotic and active tumor regions.
• ResNet-U-Net trained with T2 Fluid-Attenuated Inversion Recovery (T2-FLAIR) proved suitable for edema segmentation.
• The proposed framework effectively extracts semantic information for enhanced glioma segmentation.

Conclusions:

• ResNet-U-Net shows significant promise for automated glioma region extraction, aiding physicians in clinical decision-making.
• The study validates the practical utility and scientific grounding of the proposed deep learning framework for glioma segmentation.