MRI-based brain tumor segmentation using FPGA-accelerated neural network.

Area of Science:

• Medical Image Analysis
• Computer-Aided Detection
• Deep Learning

Background:

• Brain tumor segmentation is complex, time-consuming, and prone to errors.
• Deep learning methods improve accuracy but demand high computational resources.
• Efficient brain tumor segmentation is crucial for timely diagnosis and treatment.

Purpose of the Study:

• To develop a computer-aided detection system for efficient brain tumor segmentation.
• To accelerate the deep learning-based brain tumor segmentation process.
• To reduce computational complexity and power consumption in brain tumor segmentation.

Main Methods:

• Quantization and retraining of neural networks for brain tumor segmentation.
• Merging batch normalization layers to reduce model size and complexity.
• Designing an FPGA-based accelerator for the quantized neural network model.

Main Results:

• The FPGA accelerator demonstrated significant speed improvements over GPU and CPU.
• Achieved 5.21x and 44.47x speedup compared to TITAN V GPU and Xeon CPU, respectively.
• Showcased 11.22x and 82.33x greater energy efficiency than GPU and CPU.

Conclusions:

• FPGA accelerator enhances brain tumor segmentation speed and reduces power consumption.
• The developed system ensures high segmentation accuracy, crucial for clinical applications.
• This approach offers a new direction for automated segmentation and remote diagnosis of brain tumors.