Texture Maps as Input in 3D CNNs Applied to Classify Nodules in CT Images

Abstract

Lung cancer is the leading cause of cancer-related death worldwide. Early diagnosis of pulmonary nodules on chest CT scans provides a chance to design an effective treatment. The focus of this study is the classification problem of benign and malignant pulmonary nodules in CT images. Thus, it is proposed to apply texture maps directly to the 3D nodules as a previous of the feature extraction process. For this, the local binary patterns (LBP), with branches, such as using neighbors with borders (LBP-6), average dimensions (LBP-M), and a 3×3×3 neighborhood (LBP-3×), to highlight the nodule texture. Convolutional Neural Networks, such as DenseNet, ResNet, and LeNet, were used as attribute extractors using the 3D texture maps computed. Then, those deep features are used as input to train a Random Forest classifier. In the experiments, it is used LIDC-IDRI image database. The LIDC-IDRI database was used with two segmentation process, one made by radiologists, present in the base itself (B1), and one performed automatically by a third party (B2). In B1, the best result was the original nodules' attributes extracted with the DenseNet architecture reaching an accuracy of 0.8371, a specificity of 0.9130, sensitivity of 0.7328, and Kappa of 0.6591. In B2, the best result was a combination of attributes of the original nodule combined with the extracted LBP-6 with LeNet architecture that reached an accuracy of 0.9037, a specificity of 0.8453, sensitivity 0.9266, and Kappa of 0.7641. In conclusion, it is possible to improve the classification accuracy by including a texture map computation as part of the process.