SegGAUNet: GAN and Attention U-Net Integration for Accurate Lung Segmentation from HRCT Images

This work introduces SegGAUNet, a custom deep learning framework that merges the generative ability of Generative Adversarial Networks (GANs) with the feature-focusing mechanism of an attention U-Net for precise lung segmentation from high-resolution CT images of Interstitial Lung Disease (ILD) cases. SegGAUNet is structured as an adversarial framework with a generator and discriminator. The generator, based on attention U-Net with attention modules and skip connections, uses HRCT scans to produce binary lung masks. The discriminator assesses mask realism, providing adversarial feedback to refine predictions. Both adversarial and pixel-wise loss functions are employed, with training by checkpointing to retain the best-performing generator. A total of 217 annotated HRCT images from MedGIFT and OSIC data repositories are used, with 138 for training and 79 for testing. SegGAUNet attains 86.8% segmentation accuracy and shows superior performance compared to existing methods, even with limited data. The proposed framework demonstrates strong potential as a Computer-Aided Diagnosis (CAD) tool for ILD-affected lung segmentation, particularly in resource-limited healthcare environments.