Deep Learning-Based Multiclass Brain Tissue Segmentation in Fetal MRIs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Fetal brain tissue segmentation is crucial for identifying congenital disorders.
• Manual segmentation is time-consuming and complex due to dynamic changes in the fetal brain during gestation.
• Existing automatic methods struggle with varying tissue contrast throughout pregnancy.

Purpose of the Study:

• To develop an automated deep learning-based method for accurate fetal brain tissue segmentation.
• To reduce the manual effort required for refining segmentation results.
• To improve the quantification of fetal brain development and disorders.

Main Methods:

• A novel deep learning model incorporating a contextual transformer block (CoT-Block) in an encoder-decoder architecture.
• Integration of a hybrid dilated convolution module in the decoder to enhance feature extraction.
• Quantitative evaluation using Dice similarity coefficient (DSC), Volume Similarity (VS), and Hausdorff95 Distance (HD95).

Main Results:

• Achieved an average DSC of 83.79%, VS of 84.84%, and HD95 of 35.66 mm on 80 fetal brain MRI scans (20-35 weeks gestation).
• Demonstrated superior performance compared to existing advanced deep learning segmentation models.
• Successfully enhanced feature extraction and global contextual information capture.

Conclusions:

• The proposed deep learning method offers a significant advancement in automated fetal brain MRI segmentation.
• This technique can aid in more efficient and accurate diagnosis of fetal neurological conditions.
• The novel architectural components effectively address the complexities of segmenting the evolving fetal brain.