Abstract
In the realm of computer-aided diagnosis systems designed for lung cancer, accurately segmenting nodules holds vital importance. This segmentation process has a vital role in examining the image attributes of lung nodules captured in computed tomography scans, ultimately aiding in separation of benign and cancerous nodules. Timely detection of these lesions stands as the most effective strategy in combating lung cancer, a disease notorious for its high malignancy rates across both genders. Despite numerous deep learning techniques proposed for nodule segmentation, it remains challenging due to factors such as nodule characteristics, location, false positives, and the necessity for precise boundary detection. The present paper presents an ultra-modern method for lung nodule segmentation in computer tomographic images, based on a Generative Adversarial Network. A discriminator and a generator make up the GAN model. Our generator, Residual Dilated Attention Gate UNet, serves as the segmentation module, while a discriminator is Convolutional Neural Network classifier. To enhance training stability, we utilize the Wasserstein GAN algorithm. We compare our hybrid deep learning model, called WGAN-LUNet, both quantitatively and qualitatively with other methods that are already in use. We evaluate the model using multiple quantitative criteria.
Keywords: Deep Learning, Generative Adversarial Network (GAN), Lung nodule, Residual dilated Attention Gate UNet, Segmentation.