An efficient memory reserving-and-fading strategy for vector quantization based 3D brain segmentation and tumor...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Deep learning excels in 3D brain image segmentation, crucial for tumor diagnosis and treatment planning.
• Current methods face challenges with large manual annotation requirements and computational inefficiency.
• Radiological characteristics from tumors are vital for clinical decision-making.

Purpose of the Study:

• To develop an unsupervised 3D deep learning segmentation method for brain tumors.
• To address the limitations of manual data annotation and improve computational efficiency.
• To enhance the feasibility of deep learning models in biomedical image segmentation.

Main Methods:

• Proposed a vector quantization (VQ)-based 3D segmentation using a novel unsupervised 3D deep embedding clustering (3D-DEC) network.
• Implemented an efficiency memory reserving-and-fading strategy to boost computational speed.
• Trained the VQ-based 3D-DEC network on volume data in an unsupervised manner.

Main Results:

• Achieved superior accuracy (e.g., 91% Dice on real data) and robustness on public MRI datasets (IBSR, BrainWeb) and real clinical data.
• Demonstrated remarkable efficiency with significant speedup ratios (e.g., 31.00 on real data).
• Outperformed state-of-the-art 3D Convolutional Neural Network (CNN) models in segmentation accuracy and speed.

Conclusions:

• The proposed unsupervised method effectively addresses the lack of manual annotations in brain tumor segmentation.
• The model significantly increases computation speed while maintaining competitive segmentation accuracy.
• This approach is suitable for tumor treatment follow-ups, providing critical radiomic features for surgical and postoperative care.