GSNet: a multi-class 3D attention-based hybrid glioma segmentation network

Area of Science:

• Neuro-oncology
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Computer-aided biomedical image retrieval (CBIR) tools are increasingly vital in neuro-oncology.
• Automated brain tumor (glioma) segmentation presents challenges due to data imbalance and texture inconsistencies.
• Existing 2D methods often struggle with the complexity of 3D medical imaging data.

Purpose of the Study:

• To propose GSNet, a novel 3D convolutional neural network for robust and accurate glioma segmentation.
• To evaluate the performance of GSNet on benchmark datasets, comparing it against state-of-the-art methods.
• To demonstrate the suitability of GSNet as a reliable tool for computer-aided biomedical image retrieval in clinical settings.

Main Methods:

• GSNet, a 3D segmentation network utilizing attention-based skip connections, was developed.
• The network was trained and validated using the BraTS 2020 dataset and further tested on BraTS 2019 and BraTS 2018 datasets.
• Performance was quantified using the Dice Similarity Coefficient (DSC) for whole tumor, tumor core, and enhancing tumor segmentation.

Main Results:

• GSNet achieved high DSC scores on the BraTS 2020 dataset: 0.9239 (whole tumor), 0.9103 (tumor core), and 0.8139 (enhancing tumor).
• The model demonstrated strong performance even when trained on imbalanced datasets and showed capability in handling new data.
• GSNet outperformed several state-of-the-art methods in glioma segmentation accuracy.

Conclusions:

• GSNet is a robust and high-performing 3D network for glioma segmentation.
• The model's accuracy and reliability make it a suitable candidate for computer-aided biomedical image retrieval in neuro-oncology.
• GSNet offers a promising advancement for automated brain tumor analysis and diagnosis.