ALE-GAN: A 3D Conditional Generative Adversarial Network with Attention Learning Modules for Lung Nodule Segmentation
摘要
Early lung cancer detection using lung nodule segmentation can enhance patient survival. In computer vision applications like medical image analysis, convolutional neural networks (CNNs) outperform standard image processing methods. Convolutional neural network-based medical image segmentation methods have exhibited state-of-the-art performance, but they still face limitations. Data scarceness and class inequality induce overfitting and deprived performance. This study proposes a scalar attention learning modules (Sc-ALM) embedded generative adversarial network for lung nodule segmentation that learns nodule-focused data distribution to boost accuracy. The proposed network utilizes Sc-ALM in generator part based on U-Net as well as the discriminator based on a simple classification network that empowered the model to distinguish ground truth from falsified segmentation. Also, patch-based training reduces overfitting. The LUNA16 dataset is used to perform extensive lung nodule segmentation experiments. The proposed model’s boundary F1-score, average Jaccard index, and dice similarity coefficient are 0.9695, 0.9650, and 0.9825 respectively, outperforming various state-of-the-art models. This model has remarkable lung nodule segmentation performance and can help doctors diagnose lung nodules early by assessing their size, shape, and other properties.