Attention-Enhanced U-Net for Accurate Segmentation of COVID-19 Infected Lung Regions in CT Scans

Technical Deep Dive: Attention-Enhanced U-Net for COVID-19 Lung Segmentation

Overview

This study focused on developing a robust methodology for automatic segmentation of COVID-19 infected lung regions in CT scans. The proposed model was based on a modified U-Net architecture with attention mechanisms, data augmentation, and post-processing techniques. The approach aimed to achieve high segmentation accuracy and boundary precision.

Methodology

Dataset and Preprocessing

• The dataset was sourced from public repositories and consisted of 20 CT scans with associated lung and infection masks.
• Significant variability in Hounsfield Unit (HU) distributions was observed, prompting normalization and HU thresholding to focus on the relevant anatomical range.
• Images were resized to 128x128 pixels and annotations were binarized to 0 (background) and 1 (infected region).
• An augmentation pipeline was implemented, including rotations, translations, scaling, elastic transformations, and brightness/contrast changes, to increase dataset diversity.

Model Architecture

• The model was based on a modified U-Net architecture with a ResNet-34 backbone for feature extraction.
• Attention blocks were integrated to refine the feature maps by emphasizing infection regions and suppressing irrelevant areas.
• The decoder used transposed convolution layers for precise upsampling and segmentation map reconstruction.

Training and Optimization

• Multiple loss functions were experimented with, including Dice loss, Binary Cross-Entropy (BCE), and a hybrid BCE-Dice loss with surface loss.
• The model was trained using the Adam optimizer with a cosine annealing learning rate scheduler.
• Cross-validation and 20% of the dataset reserved for testing were used to ensure robustness.

Evaluation Metrics

• Dice coefficient, Intersection over Union (IoU), binary accuracy, mean IoU, Average Symmetric Surface Distance (ASSD), and Hausdorff distance were used to evaluate the model's performance.
• A novel generalized loss function combining weighted BCE-Dice loss and surface loss was implemented.

Results

Non-Augmented Model

• The non-augmented model achieved a Dice coefficient of 0.8502 and a mean IoU of 0.7445, with reasonable boundary alignment (ASSD 0.3907, Hausdorff 8.4853).
• However, the model lacked the ability to handle data variations, as shown by the more varied IoU distribution and lower ROC AUC of 0.91.

Augmented Model

• The augmented model significantly outperformed the non-augmented version, achieving a Dice coefficient of 0.8658 and a mean IoU of 0.8316.
• Boundary precision was also improved, with an ASSD of 0.3888 and a Hausdorff distance of 9.8995.
• The ROC AUC increased to 1.00, indicating outstanding discriminatory capability.

Interpretation

• Data augmentation and advanced architectures, such as attention mechanisms, were critical in addressing the challenges of medical image segmentation for COVID-19.
• The proposed methodology outperformed baseline models in handling variations in infection patterns and boundaries, demonstrating the benefits of the augmented dataset and architectural enhancements.

Limitations and Uncertainties

• The dataset, while publicly available, is still relatively small and may not capture the full range of COVID-19 infection patterns and imaging variability.
• The 2D segmentation approach could be extended to 3D segmentation of volumetric CT data for richer spatial context.
• Computational efficiency optimizations, such as model pruning or quantization, would be necessary for real-time clinical deployment.

Future Work

• Expanding the dataset to include more diverse cases across demographics and pathologies to improve model robustness.
• Exploring 3D segmentation techniques for volumetric CT data analysis.
• Integrating explainable AI methods to provide clinicians with insights into the model's decision-making process.
• Collaborating with interdisciplinary teams to develop comprehensive diagnostic tools for pandemic readiness.