Deep Learning Approaches for Pulmonary Nodule Classification: A CNN Algorithm for Enhanced

Abstract

For the purpose of making a preliminary diagnosis of lung cancer, computed tomography, or CT, is frequently utilized to find pulmonary nodules. However, as a result visual similarities among non-cancerous and cancerous nodules, identifying malignant from cancer nodules is not easy for doctors to accomplish. Here, a novel Convolution Neural Network architecture known as ConvNet is suggested to classify lung nodules as malignant or benign. A multi-scale, multi-path architecture is developed and utilized to increase the classification performance. This is done since there is a large variance in the nodule characteristics that are displayed in CT scan images, like Shape and Size. The multiple scale method makes use of filters of varying sizes to extract nodule features from local regions in a more efficient manner, and the multiple path architecture combines features extracted from various Convolution Network layers in order to improve the nodule features in relation to global regions. Both of these methods are part of the multi-path architecture. The LUNGx Challenge database is used to train and assess the proposed ConvNet, and it obtains specificity of 0.924, sensitivity of 0.887, and AUC of 0.948. The suggested Convolution Network is able to obtain an AUC improvement that is 14 percent higher than the current state-of-the-art unsupervised learning technique. The proposed Convolution Network also performs better than the previous state-of-the- art Convolution Networks that were specifically created for the categorization of pulmonary nodules. The suggested Convolution Networks has the potential to aid radiologists in making diagnostic judgments during CT screening when it is utilized in clinical settings.