D-UNet: A Dimension-Fusion U Shape Network for Chronic Stroke Lesion Segmentation

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Neurology

Background:

• Accurate assessment of chronic stroke lesions is vital for diagnosis, surgical planning, and prognosis.
• Convolutional Neural Networks (CNNs) show promise in medical image segmentation, but 2D CNNs miss 3D information and 3D CNNs are computationally intensive.

Purpose of the Study:

• To develop an efficient and accurate method for segmenting chronic stroke lesions.
• To address the limitations of existing 2D and 3D CNNs in medical image segmentation.

Main Methods:

• Proposed a novel dimension-fusion-UNet (D-UNet) architecture combining 2D and 3D convolutions in the encoding stage.
• Introduced an Enhance Mixing Loss (EML) function to handle data imbalance issues during network training.
• Evaluated the method on the ATLAS dataset, comparing it against three state-of-the-art techniques.

Main Results:

• The D-UNet achieved superior segmentation performance compared to 2D networks.
• The proposed architecture required significantly less computation time than 3D networks.
• The method demonstrated the best quality performance with DSC = 0.5349 ± 0.2763 and precision = 0.6331 ± 0.295.

Conclusions:

• The D-UNet presents an effective solution for chronic stroke lesion segmentation, balancing performance and computational efficiency.
• The Enhance Mixing Loss (EML) function aids in improving network training with imbalanced datasets.
• The proposed approach shows significant potential for clinical applications in stroke management.