Brain tumor segmentation based on deep learning and an attention mechanism using MRI multi-modalities brain images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Brain tumor segmentation from MRI is crucial for medical analysis but challenging.
• Existing methods using multiple MRI modalities (T1, T1c, T2, FLAIR) often have complex structures, leading to long training and testing times.
• Overfitting is a common issue in deep learning models for medical image segmentation.

Purpose of the Study:

• To develop a flexible and effective brain tumor segmentation system.
• To reduce computational time and overcome overfitting in deep learning models.
• To improve the accuracy of brain tumor segmentation compared to state-of-the-art methods.

Main Methods:

• A preprocessing approach focusing on smaller image regions to decrease computing time and prevent overfitting.
• A Cascade Convolutional Neural Network (C-CNN) designed to mine local and global features efficiently.
• Introduction of a novel Distance-Wise Attention (DWA) mechanism to enhance segmentation accuracy by considering tumor and brain center locations.

Main Results:

• The proposed C-CNN model with DWA achieved competitive results on the BRATS 2018 dataset.
• Mean Dice scores of 0.9203 (whole tumor), 0.9113 (enhancing tumor), and 0.8726 (tumor core) were obtained.
• The method demonstrated reduced computing time and overcame overfitting issues.

Conclusions:

• The proposed preprocessing and C-CNN with DWA offer an effective solution for brain tumor segmentation.
• The approach provides a balance between computational efficiency and segmentation accuracy.
• This method holds promise for improving clinical applications of brain tumor analysis.