GLAC-Unet: Global-Local Active Contour Loss with an Efficient U-Shaped Architecture for Multiclass Medical Image...

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• Traditional loss functions (Dice, Cross-Entropy) struggle with global metrics for medical image segmentation challenges like occlusion and non-uniform intensity.
• Deep learning models require effective loss functions and architectures for precise segmentation of biomedical images.

Purpose of the Study:

• To introduce a novel Global-Local Active Contour (GLAC) loss function for multiclass medical image segmentation.
• To enhance the U-Net architecture with Dense Layers, Convolutional Block Attention Modules, and DropBlock for improved feature extraction and regularization.
• To develop and validate an end-to-end trainable model (GLAC-Unet) for precise biomedical image segmentation.

Main Methods:

• Developed the Global-Local Active Contour (GLAC) loss function, integrating global and local image features within the Mumford-Shah framework for multiclass segmentation.
• Modified the U-Net architecture by incorporating Dense Layers, Convolutional Block Attention Modules, and DropBlock to enhance contextual information processing and reduce overfitting.
• Trained and evaluated the GLAC-Unet model on diverse biomedical datasets: dermoscopy (ISIC-2018), cardiac MRI (ACDC 2017), and brain MRI (Infant Brain MRI Segmentation Challenge 2019).

Main Results:

• Achieved high Dice scores: 0.9125 on ISIC-2018, 0.9260 on ACDC 2017, and 0.927 on Infant Brain MRI Segmentation Challenge 2019.
• Demonstrated statistically significant improvements (p < 0.05) over state-of-the-art models on ISIC-2018 and ACDC datasets.
• Validated the model's robustness and effectiveness across 2D and 3D medical imaging modalities.

Conclusions:

• The proposed GLAC loss function and enhanced U-Net architecture (GLAC-Unet) significantly improve multiclass medical image segmentation accuracy.
• The method effectively addresses challenges like occlusion and non-uniform intensity, outperforming existing approaches.
• GLAC-Unet shows strong potential for advancing biomedical image analysis and clinical applications.