Learning rich features with hybrid loss for brain tumor segmentation

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology

Background:

• Accurate brain tumor segmentation is crucial for diagnosis and radiotherapy planning.
• Manual segmentation is time-consuming and subjective, necessitating automated solutions.

Purpose of the Study:

• To develop an automatic and objective system for brain tumor segmentation.
• To improve the accuracy and efficiency of brain tumor segmentation in MRI images.

Main Methods:

• Proposed a parallel multi-scale feature fusing architecture with a Feature Extraction Network (FEN) and Multi-scale Feature Fusing Network (MSFFN).
• Employed two hybrid loss functions to address class imbalance issues during network optimization.

Main Results:

• Achieved Dice scores of 0.86, 0.73, and 0.61 for complete, core, and enhancing tumor regions on the BRATS 2015 dataset.
• The model has a small parameter size (6.3 MB) and outperforms state-of-the-art methods without post-processing.

Conclusions:

• The parallel architecture effectively fuses multi-level features for high-resolution segmentation.
• Hybrid loss functions mitigate class imbalance, enhancing the training process.
• The proposed method shows potential for broader applications in medical image segmentation tasks.