Study of multistep Dense U-Net-based automatic segmentation for head MRI scans.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate segmentation of head MRI scans is challenging due to equipment-dependent intensity variations.
• Existing methods struggle with precise delineation of anatomical structures.

Purpose of the Study:

• To evaluate the effectiveness of a novel multistep Dense U-Net (MDU-Net) architecture for automatic head MRI segmentation.
• To improve the accuracy of segmenting scalp, skull, and brain tissues.

Main Methods:

• A hybrid 2.5D and 3D Dense U-Net architecture (MDU-Net) was developed for initial segmentation.
• A second 2.5D Dense U-Net with histogram standardization was used for fine-grained tissue segmentation (CSF, white matter, gray matter).
• Segmentation performance was evaluated using Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and Average Symmetric Surface Distance (ASSD) on the OASIS-3 dataset.

Main Results:

• MDU-Net achieved high DSC scores across all segmented regions (scalp, skull, CSF, white matter, gray matter).
• MDU-Net demonstrated superior performance compared to Res-U-Net, Dense U-Net, U-Net++, and Swin-Unet.
• MDU-Net showed significant improvements in HD and ASSD compared to the H-Dense U-Net, particularly for gray matter segmentation.

Conclusions:

• The proposed MDU-Net model offers enhanced accuracy for automatic head MRI segmentation.
• MDU-Net outperforms existing methods in key segmentation metrics (DSC, HD, ASSD).
• This approach holds potential for improving diagnostic capabilities through more precise MRI analysis.