Optimization of deep neural network for lung cancer segmentation

Abstract

newline Lung cancer is identified as one among the most prominent kinds of tumor which affects mankind that needs innovative practice for disease prediction and treatment. Achieving a more accurate diagnosis necessitates a substantial amount of data. Yet, physicians often encounter challenges in manually analyzing extensive and intricate CT image datasets to extract essential information. Deep learning provides mechanisms for precise identification and classification of diseased tissue and its adjacent healthy structures which are vital in the diagnosis of lung cancer. This research proposes a methodology for lung cancer detection using the deep learning architecture, U-Net. newlineUNet-based architectures have demonstrated superior performance in image segmentation compared to other deep learning architectures. A U-shaped encoder and decoder are features of the selected U-Net, which is used for predicting lung areas affected by cancer from those that are not U-Net applies several convolutional and pooling layers to extract feature maps from medical images. Less segmentation accuracy is the main problem with applying the aforementioned process because of the poor resolution and lack of data in some captured medical images. Hence modified U-Net models are proposed in this study and their effectiveness is assessed. The proposed U-Net models are Modified Attention Residual U-Net (MARU-Net), Enhanced Attention Residual Parallel U-Net (EARPU-Net) and Evolutionary U-Net respectively. newlineThe Modified Attention Residual U-Net consists of ResNet and attention blocks. The attention blocks of MARU-Net separates relevant areas of the image from irrelevant areas and residual blocks helps in avoiding newlineiv newlinevanishing gradient problem in model training. The EARPU-Net model consists of four parallel U-Nets using an attention residual network, directing its attention to processing individual quadrant images rather than the entire image at once. Despite the low resolution of medical images, this strategy yields high accuracy by capturing