Attention and Residual-Atrous Convolutional Learning-Based CNN Architecture for Lung Nodule Segmentation and Classification

Different computer-aided methods serve a substantial part in the early detection of lung cancer that can significantly boost a patient's five-year survival rate. However, owing to structural resemblance, it can be difficult and time-consuming to manually distinguish malignant nodules from benign ones. Here, utilizing the notion of residual-atrous convolution and attention learning, an integrated framework has been established for nodule classification and segmentation. It can efficiently perform these tasks by seizure of comprehensively focused, high-dimensional multi-scalar features from the Computed Tomography (CT) images. In our model, the loss of crucial feature informative details that happened during transitions into deeper layers from the shallower layers is tried to be recovered through the application of original image at each convolutional learning block and inclusion of three-dimensional attention mechanisms for each decoder block. The proposed novel framework has utilized LIDR-IDRI dataset and attained the score values of 0.9520, 0.9584, and 0.9715 as Jaccard index, boundary F1-score, and average Dice similarity coefficient for the nodule segmentation task and the score values of 96.89, 95.97, and 95.84% as specificity, accuracy, and sensitivity for lung nodule classification task. The proposed automatized trainable end-to-end deep learning CNN model can assist clinicians in delivering more accurate diagnosis.