U-Grad: A Grad-CAM-Guided Reduced U-Net for Efficient Lung Cancer Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Lung cancer is a leading cause of cancer mortality globally.
• Accurate lung nodule detection from CT scans is crucial for patient management.
• Deep learning shows promise for medical image segmentation tasks.

Purpose of the Study:

• Introduce U-Grad, a novel deep learning model for lung nodule segmentation in 2D CT slices.
• Enhance nodule representation by integrating Grad-CAM heatmaps with CT images.
• Improve model interpretability and generalizability for clinical applications.

Main Methods:

• Developed U-Grad, integrating Grad-CAM heatmaps into a Reduced U-Net architecture.
• Utilized a Reduced U-Net with a maximum depth of (256,256) and Leaky Rectified Linear Unit activation.
• Trained and evaluated models on the NSCLC Radiogenomics dataset for 100 epochs.

Main Results:

• Both Reduced U-Net and U-Grad outperformed existing models.
• Reduced U-Net achieved a Dice Coefficient (DC) of 93.15% and IoU of 89.02%.
• U-Grad achieved a DC of 91.27% and IoU of 86.26%, with reduced overfitting.

Conclusions:

• U-Grad offers a robust and interpretable alternative for lung nodule segmentation.
• Explainable AI features enhance clinical utility, especially in data-scarce settings.
• The model demonstrates potential for improved lung cancer diagnosis and management.