Gradient-driven pixel connectivity convolutional neural networks classification based on U-Net lung nodule segmentation

Abstract

Lung cancer is a significant global health issue, heavily burdening healthcare systems. Early detection is crucial for improving patient outcomes. This study proposes a diagnostic aid system for the early detection and classification of lung nodules from Computed Tomography images using deep learning, based on the LUNA16 Dataset. The methodology involves three key steps. Initially, a U-Net convolutional neural network is used for semantic segmentation, followed by features extraction and selection, which are subsequently used in classification with another convolutional neural network. The segmentation using the U-Net algorithm achieved an accuracy of 99.16 % and a Dice Similarity Coefficient of 88.44 %. For distinguishing between nodules and non-nodules in regions of interest, the classification accuracy was 90.36 %. Further classification achieved 91.89 % accuracy in differentiating solid and ground glass nodules and 91.54 % in distinguishing between benign and malignant ones. These results demonstrate the model's robust performance in categorizing various nodule characteristics. These findings highlight the potential of the proposed system as a valuable tool for clinicians, contributing to improved healthcare outcomes and advancing lung cancer diagnosis and treatment.