Multi-level Kronecker Convolutional Neural Network (ML-KCNN) for Glioma Segmentation from Multi-modal MRI Volumetric...

Area of Science:

• Medical image analysis
• Artificial intelligence in radiology
• Neurosurgical oncology

Background:

• Automated glioma segmentation from MRI is challenging due to data imbalance.
• Deep learning models, particularly FCNs, advance medical image segmentation.
• Accurate segmentation aids radiologists in patient diagnosis and treatment planning.

Purpose of the Study:

• To propose a novel Multi-level Kronecker Convolutional Neural Network (MLKCNN) for automated glioma segmentation.
• To address data imbalance and improve segmentation accuracy in brain tumor imaging.
• To enhance the extraction of both local and global contextual information for precise tumor delineation.

Main Methods:

• Developed a Multi-level Kronecker Convolutional Neural Network (MLKCNN) incorporating Kronecker convolution to address missing pixels.
• Employed a Generalized Dice Loss (GDL) function to manage data imbalance issues.
• Utilized a post-processing combination of Connected Component Analysis (CCA) and Conditional Random Fields (CRF) to reduce false positives.

Main Results:

• Achieved a Dice Similarity Coefficient (DSC) of 0.74 for enhancing tumor (ET), 0.90 for whole tumor (WT), and 0.83 for tumor core (TC).
• Reduced Hausdorff Distance (HD) scores to 3.76 (ET), 4.88 (WT), and 5.85 (TC).
• Demonstrated competitive performance compared to existing brain tumor segmentation techniques through qualitative and visual evaluations.

Conclusions:

• The proposed MLKCNN effectively segments gliomas from MRI volumes, outperforming existing methods.
• The integration of Kronecker convolution, GDL, and CCA-CRF post-processing significantly improves segmentation accuracy and handles data imbalance.
• This automated system shows promise for clinical application, assisting radiologists in brain tumor diagnosis and management.