FFLUNet: Feature Fused Lightweight UNet for brain tumor segmentation.

Area of Science:

• Medical Imaging
• Neuro-oncology
• Artificial Intelligence

Background:

• Brain tumors, especially glioblastoma multiforme, pose significant challenges in neuro-oncology.
• Accurate segmentation of brain tumors in MRI scans is vital for diagnosis, treatment planning, and patient monitoring.

Purpose of the Study:

• To introduce a novel, lightweight deep convolutional neural network (CNN) model for efficient and accurate brain tumor segmentation from MRI scans.
• To develop a streamlined architecture that minimizes computational complexity while maximizing segmentation performance.

Main Methods:

• The proposed model utilizes optimized convolutional layers for capturing local and global features with minimal parameters.
• Key innovations include layerwise adaptive weighting for feature fusion, shifted windowing for improved generalization, and dynamic weighting in skip connections for balanced feature representation.

Main Results:

• The lightweight CNN model features 1.45 million parameters, representing a substantial reduction compared to existing methods like nnUNet (95% fewer) and standard UNet (91% fewer).
• Achieved a 4.9× faster GPU inference time (0.904 ± 0.002 s) compared to nnUNet (4.416 ± 0.004 s), enabling real-time deployment.
• Demonstrated competitive segmentation accuracy on publicly available MRI datasets.

Conclusions:

• The developed lightweight CNN model provides an efficient and accurate solution for brain tumor segmentation from MRI.
• Its reduced computational footprint and faster inference times make it suitable for real-time applications and deployment on resource-constrained devices.