Deep Alternate Kernel Fused Self-Attention Model-Based Lung Nodule Classification

Abstract

Lung cancer causes death with delayed diagnosis and inadequate treatment. Hence there is a need for a computer-aided detection method that can identify the nodule category whether it is benign or malignant to avoid delays in diagnosis with the help of Computerized Tomography (CT) scans. This study proposed a novel architecture Deep Alternate Kernel Fused Self-Attention Model (DAKFSAM) which utilizes the characteristics of the residual network in different forms as well as incorporates the efficiency of the attention model. This model fuses the features extracted from different alternate kernel models in three levels of process with three kinds of alternate kernel models. The self-attention model takes multiple kernel flows’ visual attention features and merges them into a form to improve nodule classification efficiency. The performance assessment utilizes the Lung Image Database Consortium-Image Database Resource Initiative (LIDC-IDRI) dataset, and the DAKFSAM mode, as proposed, achieves an F1−Score of 94.85%.