A novel end-to-end brain tumor segmentation method using improved fully convolutional networks.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Neuroimaging

Background:

• Accurate brain tumor segmentation from MRI is crucial for diagnosis and monitoring.
• Existing deep learning methods often rely on image patches, limiting contextual understanding.
• There is a need for improved end-to-end segmentation methods.

Purpose of the Study:

• To develop a novel end-to-end deep learning method for brain tumor segmentation.
• To improve information flow and feature representation within the segmentation network.
• To enable slice-by-slice automatic segmentation without relying on image patches.

Main Methods:

• A modified U-Net architecture incorporating an "up skip connection" for enhanced information flow.
• Inclusion of inception modules in each block for richer feature learning.
• A cascade training strategy for sequential segmentation of brain tumor subregions.
• An end-to-end, slice-by-slice segmentation approach.

Main Results:

• The proposed method demonstrated superior performance compared to the standard U-Net on BRATS 2015 and BRATS 2017 datasets.
• Significant improvements were observed in segmenting complete, core, and enhancing tumor regions.
• Quantitative and visual evaluations confirmed the effectiveness of the proposed modifications.

Conclusions:

• The novel end-to-end brain tumor segmentation method achieves competitive performance against state-of-the-art approaches.
• The "up skip connection" and inception modules enhance the network's ability to segment brain tumors accurately.
• The proposed method offers an effective alternative to patch-based segmentation techniques.