A Multi-Scale Context Aware Attention Model for Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image segmentation is crucial for diagnosis and treatment planning.
• Convolutional Neural Networks (CNNs), especially U-Net, excel but struggle with complex structures and variable regions of interest (ROI).
• Limitations include fixed receptive fields and information loss during down-sampling.

Purpose of the Study:

• To enhance the U-Net architecture for more consistent and accurate medical image segmentation.
• To address limitations in feature extraction and information loss in standard U-Net models.
• To introduce a novel network, Dilated Convolution and Inception blocks-based U-Net (DCI-UNet).

Main Methods:

• Modified U-Net by incorporating dilated convolution blocks for multi-scale feature extraction.
• Added a dilated inception block between encoder and decoder to reduce information recession and semantic gap.
• Integrated squeeze and excitation units to improve feature representation and alleviate vanishing gradients.
• Modified inception blocks with reduced spatial filters and dilated convolutions for larger receptive fields.

Main Results:

• The proposed DCI-UNet demonstrated improved performance on challenging medical image segmentation tasks.
• Validated on lung, skin lesion, and nucleus segmentation with varying ROI sizes.
• Outperformed existing state-of-the-art techniques, showing effectiveness and generalizability.

Conclusions:

• DCI-UNet effectively addresses limitations of standard U-Net for medical image segmentation.
• The integration of dilated convolutions and inception blocks enhances feature extraction and representation.
• The model shows significant potential for improving diagnostic accuracy and treatment planning in various medical imaging applications.