Pulmonary lesion classification using convolutional neural network for endobronchial ultrasonogram

Abstract

This dissertation aims to develop a method to help classify pulmonary lesions from endobronchial ultrasonography images by proposing new features that are extracted from an EBUS image based on medical knowledge and a pulmonary lesion classification framework. The proposed features, namely the adaptive weighted-sum of the upper triangular gray-level co-occurrence matrix and the adaptive weighted-sum of the lower triangular gray-level co-occurrence matrix are used to determine heterogeneity, which is one of the most important characteristics of malignancy. The proposed features together with other standard features are used as input data for the proposed classification framework that uses the weighted ensemble classification based on the efficacy of the other three classification models. The first model uses random forest to classify a lesion based on the combination of radiomics features, the proposed features, and the patient’s data. The second model uses EBUS images as input data to the tuned DenseNet 169 to classify a lesion. The third model uses multi-patch EBUS images as input data to the convolutional neural network to classify a lesion. The classification by the weighting ensemble function can improve the performance of the framework by combining the distinctive characteristics of each model. The data consists of 200 EBUS images from 200 patients: 124 malignant and 76 benign. From the experiment, it was found that the proposed framework yields 95% accuracy, 100% sensitivity, and 86.67% specificity.