Accurate Automatic Glioma Segmentation in Brain MRI images Based on CapsNet.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Glioma segmentation in MR images is challenging due to irregular morphology and blurred boundaries.
• Deep learning, particularly Convolutional Neural Networks (CNNs), shows promise but requires extensive annotated datasets.
• Current CNNs face limitations in data requirements for practical medical applications.

Purpose of the Study:

• To optimize a capsule neural network (SegCaps) for accurate glioma segmentation on MR images.
• To compare the performance of SegCaps against the commonly used U-Net architecture.
• To evaluate the efficacy of SegCaps in terms of accuracy, data efficiency, and parameter count.

Main Methods:

• Developed and optimized SegCaps, a capsule neural network, for glioma segmentation.
• Conducted comparative experiments using the BraTS2020 dataset.
• Trained SegCaps on 20% of the dataset and U-Net on the entire dataset.
• Utilized the Dice Similarity Coefficient (DSC) for performance evaluation.

Main Results:

• SegCaps achieved a DSC of 87.96% for glioma tumor core segmentation, outperforming U-Net's 85.56%.
• SegCaps demonstrated superior performance using only 20% of the training data compared to U-Net.
• SegCaps contains 95.4% fewer parameters than U-Net, indicating greater efficiency.

Conclusions:

• Capsule neural networks, specifically SegCaps, offer a more efficient and accurate approach to glioma segmentation in MR images.
• SegCaps' ability to generalize and learn spatial relationships makes it a promising alternative to traditional CNNs.
• The optimized SegCaps model significantly reduces data requirements and computational resources for medical image segmentation.