Pulmonary Nodule and Malignancy Classification Employing Triplanar Views and Convolutional Neural Network

Appearance of pulmonary lesions in a Computed Tomography (CT) scan of lung has the potential to become cancerous. Therefore, classification of lesion at early stage is a powerful tool to increase the chance of survival of a patient. In this paper a novel system for lesion classification as either nodule or non-nodule followed by malignancy classification (benign or malignant) is presented. Three dimensional (3D) lung tissue is constructed from the two dimensional (2D) slices of the patient CT scan. Lung is segmented by a thresholding based method, and lesions are extracted with triplanar views, axial, coronal, and sagittal. For classification, two Convolutional Neural Networks (CNNs) are cascaded, where the first network for lesion classification and the second for malignancy classification with concatenated triplaner views as an input in both stages. Experimental results performed on 1010 patients from Lung Image Database Consortium (LIDC) reveal that, the proposed approach achieved the best performance in lesion and malignancy classification with an accuracy of 82.58% and 82.62% respectively compared with existing methods.

Paper

Full text

PDF

Pulmonary Nodule and Malignancy Classification Employing Triplanar Views and Convolutional Neural Network

Semantic Scholar · Medicine · 2019

Abstract

Appearance of pulmonary lesions in a Computed Tomography (CT) scan of lung has the potential to become cancerous. Therefore, classification of lesion at early stage is a powerful tool to increase the chance of survival of a patient. In this paper a novel system for lesion classification as either nodule or non-nodule followed by malignancy classification (benign or malignant) is presented. Three dimensional (3D) lung tissue is constructed from the two dimensional (2D) slices of the patient CT scan. Lung is segmented by a thresholding based method, and lesions are extracted with triplanar views, axial, coronal, and sagittal. For classification, two Convolutional Neural Networks (CNNs) are cascaded, where the first network for lesion classification and the second for malignancy classification with concatenated triplaner views as an input in both stages. Experimental results performed on 1010 patients from Lung Image Database Consortium (LIDC) reveal that, the proposed approach achieved the best performance in lesion and malignancy classification with an accuracy of 82.58% and 82.62% respectively compared with existing methods.

Similar papers

© 2026 NYSGPT2525 LLC